Daily2026-11-14
EconomicPolicy

Huang out at Nvidia after 33 years; CFO Kress steps up interim as the board sounds Arm's Rene Haas for the permanent job.

Jensen Huang resigned as Nvidia CEO on November 13, five days after the board's independent probe opened and two days after the New York Times surfaced his second affair. CFO Colette Kress took the title on an interim basis. Financial Times reporting placed Arm chief executive Rene Haas on the board's shortlist for the permanent seat. Nvidia had participated in 105 startup investment rounds in 2026 before the resignation; one investor who has sat across the table from both Huang and Kress called her "much more conservative." Semafor separately published a Tim Teter dossier the same day covering lobbying language around chip export licenses and feature claims to Congress that did not hold up.

AI News Weekly · TechCrunch · RTÉ / FT · Semafor · 2026-11-13 · Source ↗

Nvidia's board announced Jensen Huang's resignation Friday afternoon after a week in which Fortune reported the directors had moved from probe into succession planning and the Financial Times reported they were approaching Arm chief executive Rene Haas about the permanent job. The company named Chief Financial Officer Colette Kress as interim CEO. Kress has been CFO since 2013 and runs Nvidia's investor day; the Financial Times did not put a time limit on her role, and the board did not disclose a target date for a permanent appointment. The pivot ended 33 years of founder-leadership at a company Huang took from zero-revenue 1993 startup to the only name routinely compared to Standard Oil and Bell Labs by senators asking about export controls.

What TechCrunch's ecosystem check found

Nvidia participated in 105 startup investment rounds in 2026 before Huang's resignation, per PitchBook data compiled by TechCrunch. The two largest recent checks: $800 million into Reflection AI's $2 billion Series B, and the pending $13 billion acquisition of Hugging Face, expected to close in 2027. Founders talking to TechCrunch set the tone. "Nvidia is carrying almost every AI-native startup — applications, neoclouds, inference, foundation models," said Shamillah Bankiya of Dawn Capital. An unnamed investor who had been on the other side of multiple Nvidia-led rounds told TechCrunch the shift to Kress is a shift in risk appetite, saying she is "much more conservative." Startups backed by Nvidia have raised $100-to-$500 million for GPU purchase commitments at $1 billion+ valuations; Crusoe's $1.4 billion Series E is one example named, with CoreWeave, Mistral AI, Lambda and Black Forest Labs cited alongside. Founders now expect smaller Nvidia checks, fewer new logo investments, and lower valuations on followup rounds.

"Nvidia is carrying almost every AI-native startup — applications, neoclouds, inference, foundation models." — Shamillah Bankiya, Dawn Capital partner, to TechCrunch, 2026-11-13

The Kress question

The CFO's role at Nvidia for the last three years has been less about the balance sheet than about keeping buyside orderly through a GPU capex cycle nobody has a precedent for. Kress has done that job. The question raised by the ecosystem reaction is whether she carries the same instinct for writing checks into companies that do not yet exist as a line item. The Wall Street Journal reported on Friday that at an August chip-backstop meeting Huang referred to Kress as "a deer in the headlights"; the WSJ framed the remark as a private aside rather than a formal assessment, but the comment put a floor under the interim framing. The AI Weekly aggregation of TechCrunch's piece went up at 23:11 UTC on November 13; TechCrunch's primary byline is Rebecca Bellan, with reporting support from Kirsten Korosec.

"A deer in the headlights." — Jensen Huang on Colette Kress, August 2026 chip-backstop meeting, per Wall Street Journal

Haas on the shortlist

The Financial Times reported, in reporting cited by RTÉ, that Nvidia's board has approached Arm chief executive Rene Haas about the permanent CEO role. Haas has run Arm since February 2022, taking it through the September 2023 Nasdaq IPO and overseeing a 2026 product cycle in which Arm-based server designs reached parity with x86 on AI-training workloads in independent benchmarks. The approach is early-stage. Arm declined to comment to the Financial Times. The Haas candidacy carries one interesting feature for Nvidia's export-control arc: Arm sits at the center of the CPU ecosystem that competes directly with Intel on the host side of every Nvidia GPU server, and Haas has been the public face of the industry push against the Trump administration's broader chip restrictions.

The Semafor file

Semafor published an unrelated-but-related dossier the same day on Nvidia general counsel Tim Teter. In reporting by Reed Albergotti, Semafor said Teter told government officials in January that "the president wanted" chip export licensing conditions to move faster, and that Nvidia opposed a know-your-customer threshold on export licenses that the company's counsel internally described as a "poison pill." The piece added that Nvidia lobbyists told Congress location tracking on exported chips was technically infeasible; Nvidia then demonstrated and launched that feature within months. The Semafor piece connects the lobbying arc to two sales findings already in the public record: Bloomberg reported Nvidia undercounted sales to Megaspeed by 50,000 units, and The Information put the dollar value of chips sold to entities linked to the alleged Chinese asset at $7.5 billion. The Semafor piece landed roughly five hours before the resignation confirmation.

"The president wanted it." — Tim Teter, Nvidia general counsel, to government officials on export-license pace, January 2026, per Semafor

What the board has not said

Nvidia's November 13 statement did not disclose the compensation terms of Huang's departure, whether his 3.5-percent personal stake carries continued board representation, or whether the November 17 third-quarter earnings call goes forward with Kress on the line. Three House Select Committee on China briefings scheduled for November 18 remain on the calendar; a spokesperson for chair John Moolenaar told Semafor the committee "has not been formally informed" whether Huang or Kress will appear. The Shenzhen US-China "superintelligence dialogue" Treasury Secretary Scott Bessent organized for the APEC margins on November 18-19 was built around Huang; the State Department has not said whether Kress or a second executive attends. The AI chip industry opened Friday afternoon with Nvidia stock down about 9 percent in late trading, extending the arc that started with the November 8 board probe statement.

Big Picture

Jensen Huang resigned from Nvidia on Friday. Colette Kress is interim CEO; the Financial Times reported the board has approached Arm chief executive Rene Haas about the permanent seat. TechCrunch tallied 105 Nvidia startup investment rounds in 2026 before Friday, including an $800 million check into Reflection AI and a pending $13 billion acquisition of Hugging Face. Shamillah Bankiya of Dawn Capital told TechCrunch Nvidia is "carrying almost every AI-native startup — applications, neoclouds, inference, foundation models." An unnamed investor called Kress "much more conservative"; the Wall Street Journal reported Huang called her "a deer in the headlights" at an August chip-backstop meeting. Semafor's Reed Albergotti published a Tim Teter dossier the same day covering "the president wanted" licensing pressure, an internal "poison pill" tag for a KYC export threshold, and lobbyist claims to Congress that chip location tracking was infeasible months before Nvidia launched the feature. Bloomberg previously reported a 50,000-unit Megaspeed undercount; The Information put sales to entities linked to the alleged Chinese asset at $7.5 billion.

The FAA grounded every departing US flight Friday, the first full nationwide ground stop since January 2023, three days after it rolled Air Space Intelligence's SMART Strategic Management of Airspace Routes and Trajectories tool from a Washington-only pilot to all 22 regional air route traffic control centers on a 12-year, $875 million contract that ASI won over Palantir and Thales. Transportation Secretary Sean Duffy blamed "a communication issue impacting all Air Route Traffic Control Centers" and did not mention SMART; Homeland Security administrator Markwayne Mullin told a reporter the agency was "looking at that." ASI denied its product or integration was involved. The question Futurism and AI Weekly flagged: whether the FAA has independent instrumentation to distinguish a SMART failure from a comms outage, or whether ASI's logs end up as the official record.

The frameworks layer hardened. llama.cpp and ggml shipped v1.0.0 after 3.5 years with a public semantic-versioning contract: major bumps for breaking libllama changes, minor for new backends and models, patch for fixes. Six Hugging Face authors are credited with 104 community co-contributors; NVIDIA, AMD, Qualcomm, Arm, IBM and Intel supply self-hosted CI runners. Perplexity released Decider v1.3 and Flash v1.3 on Hugging Face under Apache 2.0 at 91.6 percent average across 63 query-classification tasks, 93.98 percent agreement with human judges across 2.3 million questions from a million real requests, and $26.47 per million production requests ($13.24 for Flash); the models start from Qwen weights with bidirectional attention. OpenAI dissolved its standalone robotics team for the second time in six years, with research moving under Chief Research Officer Mark Chen and applied under President Greg Brockman; researchers argued general models can run robots directly, citing Nvidia's RoboLab-120 benchmark where GPT-6 Astra posted 74.8 percent success against 51.0 for specialized robotics models. Sam Altman still targets a first public humanoid demo in 2027.

Money kept moving despite the Nvidia transition. LangChain closed $300 million Series C at a $3 billion valuation led by CapitalG with GIC, DFJ Growth, IVP and Sapphire Ventures participating, with LangSmith at 7,000 customers including Nvidia, Toyota, Salesforce, Harvey, Rippling and Clay. Moonshot AI deactivated its Cayman holdco and shifted investor stakes to a mainland China operating entity under tighter China Securities Regulatory Commission rules; the Hong Kong Stock Exchange application targets a Q1 2027 offering of up to $5 billion at a $50 billion valuation, up from a $31.5 billion summer round, with ARR "currently at $1 billion and on track to double to $2 billion by December" per the reporting. Anthropic's amended S-1 set the ANT price band at $70 to $78 per share for 825 million shares with pricing November 18 and trading November 19, implying up to $1.4 trillion at the top of the range against an $80 billion first-half loss. OpenAI launched ChatGPT Wallet for Pro users 18+ in the US with American Express, Mastercard and Visa network-tokenized cards and Face-ID or fingerprint confirmation; Visa's internal testing reportedly found agents fail 42.8 percent of transactions. Jan von Plato used Claude to catalogue 663 theorems from Gödel's 1,009-page Gabelsberger-shorthand notebooks including a near-complete reconstruction of the independence of the axiom of choice that Gödel claimed in 1966 and Paul Cohen later won a Fields Medal for; Coq verified 400 as stated, 200 needed sharper statements, 50 were false. Hugging Face Daily Papers filled with 32 entries led by NUS SO-Arena at 73 upvotes, NVIDIA BirdDreams at 46, DAGroup PhysLoop at 42, LongCat G-STAR at 40 and inclusionAI Duo at 35. On policy, EU Advocate General Maciej Szpunar issued an opinion November 3 in Like Company v. Google Ireland that chatbots reusing news content to generate answers may infringe EU press publisher rights, with the opinion landing on the Court of Justice's docket for a binding decision in 2027.

The rundown

PolicySafety

FAA grounds every US departing flight three days after Air Space Intelligence's SMART tool went nationwide; DHS says it is "looking at" the system.

Futurism · AI News Weekly · 2026-11-13 · Source ↗

The first full nationwide ground stop since January 2023 arrived three days after the FAA rolled SMART to all 22 regional air route traffic control centers. Transportation Secretary Sean Duffy blamed "a communication issue"; DHS administrator Markwayne Mullin said on camera the agency is "looking at that." ASI, which won the 12-year $875 million contract in June over Palantir and Thales, denied the stoppage was related to SMART.

Every departing US flight stopped Friday afternoon in a nationwide ground stop that Futurism reported was the first of its kind since January 2023. The timing is the story: three days earlier, the FAA finished rolling Strategic Management of Airspace Routes and Trajectories (SMART) from a Washington, DC pilot to all 22 regional air route traffic control centers. SMART, built and operated by startup Air Space Intelligence (ASI), is marketed to the FAA as a shift from reactive to predictive air traffic management; the public pitch is fewer delays, less fuel burn, and smoother passenger experience.

Who said what

Transportation Secretary Sean Duffy, in the televised remarks after the stop, blamed "a communication issue impacting all Air Route Traffic Control Centers" and made no mention of SMART. Homeland Security administrator Markwayne Mullin, pressed on camera about SMART by a reporter, said the agency was "looking at that" and was coordinating with other agencies on a possible link. ASI issued a Friday-evening statement saying the stoppage was not related to SMART, any of its products, or its FAA integration. Neither the FAA nor Secretary Duffy has since publicly linked SMART to the stop or cleared it.

"We're looking at that. We're working with other agencies." — Markwayne Mullin, DHS administrator, on SMART, 2026-11-13

The contract and the vendor

ASI won the 12-year, $875 million FAA SMART contract in June 2026, beating Palantir and Thales. The June announcement framed the award as the FAA's first production AI deployment at nationwide scale. Prior to Friday the system operated only in the Washington, DC regional center. The rollout to the other 21 centers was completed Tuesday morning per FAA release notes. Futurism's dispatch flagged the three-day window between the national rollout and the ground stop as "the central public question of the week"; ASI's denial, written as a bounded denial about its infrastructure rather than a diagnosis of the failure, leaves the diagnosis itself open.

"The stoppage was not related to SMART or any of our products or integration with the FAA." — Air Space Intelligence statement, 2026-11-13

The attribution question

AI News Weekly's editorial note flagged the structural problem for the FAA: whether the agency has independent instrumentation to distinguish a SMART failure from a comms outage, or whether its post-mortem relies on ASI's own logs. If the vendor's logs become the official record, the vendor's denial effectively settles the record. The FAA has not disclosed a technical readout timeline. Secretary Duffy's wording, which neither named nor cleared SMART, was framed by AI News Weekly as the first public dispute over attribution of blame for AI in safety-critical federal infrastructure.

Futurism also noted the political overlay: the SMART contract award over Palantir came after a public lobbying push by Palantir for the role, and the FAA's rollout timetable was accelerated under direct pressure from the Trump administration's "Golden Age of US Aviation" initiative announced in May. Palantir declined to comment to Futurism on Friday.

Economic

LangChain closes $300M Series C at $3B led by CapitalG; LangSmith hits 7,000 customers including Nvidia, Toyota, Salesforce and Harvey.

TechCrunch · 2026-11-13 · Source ↗

GIC, DFJ Growth, IVP and Sapphire Ventures joined. The $3 billion valuation more than doubles LangChain's $1.25 billion mark from roughly a year earlier when it raised $125 million. TechCrunch frames the round as enterprises hedging against single-model lock-in, with LangSmith's "vendor-neutral" posture the selling point against first-party agent frameworks from Anthropic, OpenAI and Google.

LangChain closed a $300 million Series C at a $3 billion valuation led by Alphabet-backed CapitalG with participation from Singapore sovereign fund GIC, DFJ Growth, IVP and Sapphire Ventures. TechCrunch reported the round on Friday. The $3 billion mark more than doubles the $1.25 billion LangChain carried after its Series B roughly a year earlier, when it took $125 million from the same tier of growth investors. LangSmith, the company's agent-development and observability platform, reports 7,000 customers.

What the customer list says

TechCrunch named Nvidia, Toyota, Salesforce, Harvey, Rippling and Clay among LangSmith's enterprise users. The set spans chipmaker, carmaker, enterprise SaaS, legal AI, HR and go-to-market tooling. LangChain positions LangSmith as vendor-neutral, meaning the stack does not tie customers to a specific model provider, cloud, or ecosystem. That posture matters now because every frontier lab has its own first-party agent framework: Anthropic's Claude Agents, OpenAI's Dots, Google's Gemini Agent Builder, Meta's Muse Agents. Enterprises rolling out agents across departments want a control plane that outlives any single provider's roadmap. The round reads as a bet that this preference is durable.

"7,000 customers and counting — and we did it while staying vendor-neutral." — LangChain spokesperson paraphrased in TechCrunch coverage, 2026-11-13

The context TechCrunch did not quite name

LangChain started in 2022 as an open-source project aimed at early LLM limits around web search, API calls, and database interaction. As model capability improved, the product moved up a layer to manage agents rather than compensate for model gaps. LangSmith is now the commercial anchor; TechCrunch did not split the 7,000 customer count into paid and free users, did not disclose revenue, and did not report net-retention. Those omissions matter for a growth-stage company at a 24x step-up on roughly 12 months, but TechCrunch did report that AI News Weekly has tracked 441 funding alerts across the AI stack in the last 90 days, a figure that frames how dense the round calendar is.

The CapitalG read

CapitalG has been consistently early into the picks-and-shovels tier of the AI stack. The firm led LangChain's Series A in late 2023 and the company's Series B in late 2025; leading the C locks the Alphabet growth arm to the full capital structure through a plausible IPO window. GIC's participation adds Asian institutional capital on the back of Singapore's sovereign-fund rotation into enterprise AI infrastructure. DFJ Growth, IVP and Sapphire Ventures round out the US growth-equity bench.

TechEconomic

OpenAI dissolves standalone robotics team for second time in six years; research moves under Mark Chen, applied under Greg Brockman.

Fortune · 2026-11-13 · Source ↗

Researchers now report to Chief Research Officer Mark Chen; applied robotics and hardware staff report to President Greg Brockman. The stated rationale: general models such as GPT-6 Astra can control robots directly, citing a Nvidia RoboLab-120 success rate of 74.8 percent against 51.0 for specialized robotics models. Sam Altman still targets a first public humanoid demonstration in 2027. At CoRL in Austin this week OpenAI's booth sat empty despite being registered.

OpenAI has dissolved its standalone robotics team, Fortune reported Friday. It is the second time in six years that the company has wound down a dedicated robotics unit. The first closure in 2020 saw cofounder Wojciech Zaremba cite "limited robot data" as the reason the specialized effort could not pay for its own GPU cost. OpenAI rebuilt the team in 2024 around a partnership with humanoid startup Figure; that partnership ended in February 2025, well before Figure's December 2025 pivot to Brett Adcock's own foundation model. The rebuilt unit survived the Figure breakup; what killed it this week is a research-side argument that it is no longer necessary.

Where the people go

Research staff from the robotics team now report to Chief Research Officer Mark Chen. Applied robotics and hardware staff report to President Greg Brockman. Fortune's reporting did not disclose total headcount, and OpenAI did not confirm whether anyone was laid off in the transition. A spokesperson told Fortune robotics "remains a core part of OpenAI's long-term vision" and described the move as acceleration rather than retreat. CEO Sam Altman is still targeting a first public demonstration of OpenAI's own humanoid robot in 2027.

"Robotics remains a core part of OpenAI's long-term vision." — OpenAI spokesperson to Fortune, 2026-11-13

The benchmark that got cited

Fortune reported that the research argument inside OpenAI was built on Nvidia's RoboLab-120 benchmark, where GPT-6 Astra reached a 74.8 percent success rate on a simulated manipulation benchmark against 51.0 percent for specialized robotics models. The gap matters because it suggests that the part of the robotics stack worth specialized investment, the part the 2024 rebuild existed to produce, is thinner than previously believed; general-purpose multimodal models with action heads trained on the same data are close enough that the specialized stack carries a marginal cost. The benchmark is a simulated one. Fortune flagged the open question itself: whether outsiders can reproduce the 74.8-versus-51.0 result on real hardware.

What the outside robotics community said

Guanya Shi of Carnegie Mellon Robotics Institute told Fortune the move suggests OpenAI believes robotics is "no longer a special problem" for the lab, and that the shift tracks a broader thesis that generalist foundation models absorb robotics as a capability rather than a discipline. Russ Tedrake of MIT, who runs Walden Robotics, pushed back: in his framing, robotics startups still hold an edge from real-world deployment experience that simulation benchmarks under-value. Pete Florence, Generalist cofounder, told Fortune it is "unclear that any existing model would dominate robotics," framing the specialized-versus-generalist split as an empirical question still open.

"Robotics is no longer a special problem for OpenAI." — Guanya Shi, Carnegie Mellon Robotics Institute, to Fortune

The empty booth in Austin

At the CoRL robotics conference in Austin this week, OpenAI's booth sat empty despite being registered. Fortune noted the detail without comment. For the specialized-robotics startup ecosystem that has raised against the premise OpenAI intends to be a significant hardware buyer or research partner, Friday's news is the second-largest demand-side signal of the week, after the Nvidia transition. Investors at CoRL told Fortune they expect the valuation reset on humanoid-robotics startups to be sharper than the one already underway in AI infrastructure.

Tech

Perplexity ships Decider v1.3 and Flash v1.3 on Hugging Face at 91.6 percent average across 63 classification tasks and 83.99 nDCG@10 on retrieval.

Perplexity blog · Hugging Face · 2026-11-12 · Source ↗

Both models ship under Apache 2.0 with Decider v1.3 at 91.6 percent across 63 query-classification tasks, Flash v1.3 priced at $13.24 per million against $26.47 for v1.3. Retrieval benchmark: 83.99 nDCG@10 across 72 tasks. 93.98 percent agreement with human judges across 2.3 million questions drawn from a million real requests. Starts from Qwen weights with attention layers made bidirectional.

Perplexity released Decider v1.3 and the smaller Decider Flash v1.3 Thursday under Apache 2.0 on Hugging Face. The models are the latest in a Decider line that now spans v1, v1.1, v1.2, Flash v1.2 and v1.3. Perplexity positions them as successors to those prior versions; the lineage matters because Decider has quietly become the control-plane choice for teams who want to route or classify LLM traffic without burning full generative-model tokens on every call.

The headline numbers

Decider v1.3 averages 91.6 percent accuracy across 63 query-classification tasks. Perplexity says that is 6.54 points above its production classifiers. The mean across eight languages lands at 88.16 percent. On a separate retrieval and reranking benchmark spanning 72 tasks it scores 83.99 nDCG@10. On human-agreement, Perplexity reports 93.98 percent agreement with human judges across 2.3 million questions drawn from a million real production requests. In seven domain sweeps, Perplexity reports leads of up to 12.9 points in finance and 9.3 in law against a competing model it calls Jev 1.14.

"Across 2.3 million questions from one million real requests, v1.3 agrees with the human judge 93.98 percent of the time." — Perplexity Decider v1.3 release, 2026-11-12

The architecture choice

The models start from Qwen generative weights. The core modification: Perplexity makes the attention layers bidirectional, so each token can use context from the whole input. Decision models of this family return yes/no, multiple-choice, or ordered-score answers rather than open-ended text; they are optimized for the routing and classification layer that feeds every production RAG, agent, and tool-use system. A new "Locate" feature extends the model to extract text spans, taking it past pure yes/no routing into light extraction territory.

The pricing arithmetic

A million production requests cost $26.47 on Decider v1.3 and $13.24 on Flash. Availability is across Hugging Face, OpenRouter and the Perplexity Decisions API. Perplexity positions the release against Google Decision 1, OpenAI GPT-6 Luna and Amazon Strands Decider 2B. One internal datapoint Perplexity volunteered: the model is being used at 1 billion tokens across 7,000 agent sessions in one day inside the company. The benchmark claims are Perplexity's own; the Apache 2.0 release means independent reproduction is straightforward, which is likely how the production line gets stress-tested over the coming weeks.

"Each token can use context from the whole input." — Perplexity release note on the bidirectional attention modification
ApplicationSafety

OpenAI launches ChatGPT Wallet so "dots" agents can buy for Pro users after a Face-ID confirm; Visa testing put agent-transaction failure at 42.8 percent.

CNBC · 2026-11-13 · Source ↗

Available first to a limited group of US Pro subscribers aged 18 and up on American Express, Mastercard or Visa network-tokenized cards. Neither the agent nor the merchant sees the real card number; each purchase uses a network-issued token with its own limited-use number, security code and expiration. OpenAI collects no share of transaction revenue at launch. Amazon reportedly blocks AI-agent purchases to protect Alexa.

OpenAI launched ChatGPT Wallet Friday so its "dots" agent can make purchases on behalf of the user after a one-time card entry. The feature is restricted at launch to a limited group of US Pro subscribers aged 18 and up, on American Express, Mastercard or Visa cards issued in the United States. The product sits between agent and merchant: the agent assembles the order, the user is presented with items, total and card, then confirms with a biometric check such as Face ID or a fingerprint before the charge posts.

The security architecture

Neither the agent nor the merchant ever sees the real card primary account number. Each purchase uses a network-issued token with its own limited-use number, security code and expiration date. Cardholders keep their existing rewards and protections; the biometric confirm step is non-bypassable. CNBC reported OpenAI is not collecting any share of transaction revenue at launch. The company separately published a Commerce Partner Playbook aimed at retailers who want better placement inside dots conversations. OpenAI did not disclose whether Playbook participation translates to ranking preference visible to the user.

"Each transaction uses a network-issued token with its own limited-use number, security code and expiration — merchants never see the primary account number." — OpenAI ChatGPT Wallet FAQ, 2026-11-13

What the testing numbers say

Visa's internal testing of agent-driven payments reportedly found agents failed 42.8 percent of transactions, a number CNBC flagged as a sign merchant infrastructure is not yet ready to absorb agent flow. The common failure modes were address mismatches, SKU drift between agent-displayed and checkout-rendered items, and out-of-stock events the agent did not detect before confirmation. Amazon reportedly blocks AI-agent purchases through its site to protect the Alexa commerce path. Visa did not comment on record to CNBC.

"Agents failed 42.8 percent of transactions in our testing." — Visa internal testing results, per CNBC reporting

The open questions

CNBC's reporting did not resolve two consumer-protection questions: which dispute process applies when an agent buys the wrong item (cardholder chargeback, merchant policy, or a new OpenAI-arbitrated flow) and whether retailers who participate in the Commerce Partner Playbook get ranking preference the user is told about. The Playbook itself is not public; OpenAI has provided it to a bounded set of US retailers. A separate question raised by the launch: whether the network-token architecture scales to card-not-present flows outside OpenAI's control, or whether merchants have to adopt agent-ready checkout endpoints. Neither Visa nor Mastercard disclosed network-wide agent-ready merchant counts.

ApplicationTech

Jan von Plato uses Claude to surface 663 theorems from Gödel's 1,009-page Gabelsberger-shorthand notebooks — including a Coq-verified independence-of-choice reconstruction.

The Guardian · 2026-11-13 · Source ↗

Finnish logician Jan von Plato at the University of Helsinki has worked on Gödel's notebooks since 2017 at roughly a page per week. In September he gave Claude his transcriptions and six of his own books on Gödel; within days Claude returned a 650-page explanation and a 663-theorem catalogue. Von Plato checked with Coq: ~400 verified as stated, ~200 needed sharper statements, ~50 were false. Headline find: a Coq-verified reconstruction of the independence of the axiom of choice Gödel claimed in 1966 and Paul Cohen won a Fields Medal for.

The Guardian published a dispatch from Helsinki Friday on what is the most consequential formal-reasoning application of a frontier model this quarter. Jan von Plato, the Finnish logician at the University of Helsinki, has worked since 2017 on the Nachlass of Kurt Gödel, 1,009 pages of notebooks written in Gabelsberger, a German shorthand already obsolete in Gödel's own lifetime. The pace had been roughly one page per week.

The method

In September 2026, von Plato gave Claude his accumulated transcriptions and six of his own books on Gödel's work. Within days, Claude returned a 650-page explanation of the notebooks and a catalogue of 663 theorems Gödel proved, sketched or outlined across the notebooks. Von Plato then ran the catalogue through Coq, the standard formal-proof verifier. Coq validated roughly 400 theorems as stated. Roughly 200 required sharper statements or additional assumptions to check. Roughly 50 were false. Gödel's own marginal annotations, including "falsch" (incorrect) and "schlecht" (bad), fit the error pattern.

"It is a task basically impossible for a human to do." — Jan von Plato, University of Helsinki, 2026-11-13

The headline find

The Guardian's focus is on a near-complete reconstruction of the proof of the independence of the axiom of choice inside the notebooks. Gödel claimed the result in 1966 but never published it. Paul Cohen later won the Fields Medal for proving the same result. Claude's reconstructed version of Gödel's draft argument passed Coq verification. If the reconstruction holds under independent review, it reassigns the credit timeline of one of the twentieth century's most-cited results in mathematical logic. Von Plato's framing: the notebooks also predict results that only became known in the 1970s and 1980s, and they contain 14 further logic topics that remain undeveloped in subsequent literature.

"Separating Claude's contribution from Gödel's is a job for the historians." — Joel David Hamkins, Notre Dame

The caveats

Joel David Hamkins of Notre Dame told The Guardian that disentangling what Claude added from what Gödel put on the page is "a job for the historians" — a bounded acknowledgement that the 650-page explanation is a collaboration whose credit boundary is not yet fixed. The axiom-of-choice reconstruction still awaits independent logician review. No data release has yet enabled outside replication; von Plato's protocol is privately archived. The Guardian's open question, which AI Weekly's editorial note echoes: how the ~200 "needed sharper statements" entries divide between Claude's phrasing and von Plato's own interpretive repairs.

Tech

llama.cpp and ggml ship v1.0.0 after 3.5 years with a public semver contract; Nvidia, AMD, Qualcomm, Arm, IBM and Intel contribute self-hosted CI.

ggml-org · Hugging Face blog · 2026-11-13 · Source ↗

Both libraries now follow semantic versioning: major bumps mark breaking changes to the public C API in `./include`; minor bumps mark backward-compatible features, new backends, new model support; patch bumps mark fixes without API change. llama.cpp supports more than 150 architecture variants and all mainstream speculative decoding methods. Six Hugging Face authors are credited with 104 community co-contributors.

ggml-org released llama.cpp v1.0.0 and ggml v1.0.0 on Friday, three and a half years after Georgi Gerganov started the project. The release is a versioning commitment rather than an API shake-up: both libraries now carry an explicit semantic-versioning contract. Major bumps mark breaking changes to the public C API in `./include`. Minor bumps mark backward-compatible features, new backends, new model support. Patch bumps mark bug fixes that do not change the API. The team said both projects will signal API and ABI incompatibilities through major-version bumps and will make those boundaries clearer to downstream users as adoption grows.

Scope

llama.cpp supports more than 150 architecture variants at the 1.0 mark, including every mainstream speculative decoding method. Pre-release artifacts cover Linux, macOS, Windows, Android and iOS. Packaging is available through Docker, Debian, Homebrew, conda-forge, MacPorts, Winget and an XCFramework for iOS/macOS. Self-hosted CI runners are contributed by Nvidia, AMD, Qualcomm, Arm, IBM and Intel; scaling uses Hugging Face infrastructure such as HF Buckets and HF Jobs. Six Hugging Face authors are credited by handle on the release: julien-c, ggerganov, clem, pcuenq, raunoryys, nisbenz. They are joined by 104 additional community co-contributors.

"The most lightweight, versatile and sensible platform for building cross-platform and private AI products." — ggml-org on llama.cpp v1.0.0, 2026-11-13

What the semver commitment buys downstream

The practical implication for the dozens of commercial inference products built on libllama — Ollama, LM Studio, the various cloud providers' on-device SDKs, and the handful of enterprise on-prem inference stacks that have libllama at their core — is a contract about breakage that did not previously exist. Downstream maintainers now know which releases to pin on which trajectory, which bump demands a code change, and which does not. For enterprise and distribution use, where the dependency-stability bar is higher than a fast-moving research codebase typically supports, the semver commitment is the headline change. The release notes do not enumerate specific API changes or deprecations at the 1.0 boundary; the Hugging Face blog post is the place to check for the detail.

The quiet alignment story

The six-vendor CI contribution list is the other noteworthy detail. Nvidia, AMD, Qualcomm, Arm, IBM and Intel each run self-hosted runners for the project; on a day when Nvidia's CEO resigned and Qualcomm's AI200/AI250 cards were still working their way through hyperscaler qualification, the fact that all six carry llama.cpp as a shared primitive is a signal that the on-device and edge-inference layer has consolidated around the project in a way the server-inference layer has not yet consolidated around any one framework.

Economic

Moonshot AI unwinds Cayman holdco and shifts investor stakes to mainland China entity for a $50B Hong Kong IPO targeting Q1 2027.

Wall Street Journal · SCMP · 2026-11-13 · Source ↗

The deactivated Cayman vehicle fully unwinds the variable-interest-entity structure, driven by tighter China Securities Regulatory Commission rules on offshore listing shells. The confidential Hong Kong Stock Exchange application targets a Q1 2027 offering of up to $5 billion at a $50 billion valuation, up from a $31.5 billion summer round. ARR "currently at $1 billion and on track to double to $2 billion by December." A data-security probe into Moonshot and DeepSeek remains open.

Moonshot AI has deactivated its Cayman Islands holding company and is moving investor stakes to its mainland China operating entity. The Wall Street Journal report, with South China Morning Post confirmation, frames this as the full unwinding of the VIE (variable-interest-entity) structure that has been the standard vehicle for Chinese technology listings outside mainland China for the last two decades. The driver is tightened China Securities Regulatory Commission rules on offshore listing vehicles. The commercial consequence is a confidential Hong Kong Stock Exchange application targeting a Q1 2027 offering of up to $5 billion.

The valuation arithmetic

Moonshot's last private round this summer landed at $31.5 billion. The new valuation implied by the HKEx application is approximately $50 billion, a 58-percent step-up over roughly five months. Reported annual recurring revenue is "currently at $1 billion and on track to double to $2 billion by December," per WSJ's reporting of Moonshot's own financial disclosure. At $50 billion the implied forward revenue multiple is roughly 25x on the $2 billion target, in line with the multiples Anthropic and OpenAI are commanding in private secondaries and below the multiple implied by Anthropic's Nasdaq IPO price band.

"Annual recurring revenue is currently at $1 billion and on track to double to $2 billion by December." — Moonshot AI financial disclosure, per Wall Street Journal

The regulatory overhang

Chinese regulators have opened a data-security probe into Moonshot and rival DeepSeek. Moonshot executives are also reportedly trying to settle an investor dispute before any listing. The WSJ framed the data-security probe, not the offshore unwinding, as the primary threat to the $50 billion valuation. An unresolved question flagged by AI Weekly's editorial note: the price at which the dollar-denominated Cayman-structured early investors are being cashed out, and whether the step-down from an offshore-friendly to an onshore structure carries a drag the public prospectus will have to disclose.

The Anthropic parallel

The WSJ piece noted a separate report that Anthropic's IPO underwriters have excluded China and Hong Kong investors from the allocation. Moonshot's move the other direction — unwinding the Cayman structure to list in Hong Kong — tracks the hardening of the US-China capital-markets barrier at the AI-company scale. The question for Q1 2027 is whether Hong Kong investor demand absorbs a $5 billion IPO at the proposed valuation without a USD institutional base; Hong Kong's recent AI IPO issuance has been thinner than the HKEx post-restructuring pipeline implied.

Policy

Semafor: Nvidia counsel Tim Teter told officials "the president wanted" licenses to move faster; opposed a KYC threshold the company tagged internally as a "poison pill."

Semafor · 2026-11-13 · Source ↗

Reed Albergotti's Semafor piece connects January lobbying language around chip export licensing to documented sales findings: Bloomberg reporting on an undercount of 50,000 Megaspeed units, and The Information's $7.5 billion figure for chips sold to entities linked to the alleged Chinese asset. Nvidia lobbyists told Congress location tracking was infeasible; Nvidia later demonstrated and launched the feature within months.

Semafor's Reed Albergotti published a dossier Friday afternoon on the lobbying pattern inside Nvidia's chip-export-license operation, with general counsel Tim Teter at the center. The piece anchors on one line: in January, Teter told government officials that "the president wanted" chip export licensing conditions to move faster. The claim appears to have been made in a context where Nvidia's internal position favored a loosening of a proposed know-your-customer threshold the company's own counsel described internally as a "poison pill."

The operational claim versus the operational deliverable

Semafor documents a second inconsistency. Nvidia lobbyists told Congress that location tracking on exported chips was technically infeasible. Nvidia subsequently demonstrated and launched that feature in a product release months later. The piece frames the gap between the lobbying claim and the delivered capability as a pattern rather than a one-off, and names a second case where Nvidia counsel's representations about enforcement costs differed from subsequent product moves.

"The president wanted it." — Tim Teter, Nvidia general counsel, to government officials on export-license pace, January 2026

What the sales record shows

Semafor connects the lobbying arc to two sales findings in the public record. Bloomberg reported that Nvidia undercounted sales to Megaspeed by 50,000 units; that figure came out of a financial-review process Bloomberg described in detail. The Information put the dollar value of chips sold to entities linked to the alleged Chinese asset at $7.5 billion. Albergotti's piece does not resolve the question of what Teter personally knew, when he knew it, or whether the internal "poison pill" framing was a legal judgment or a commercial one. It does, however, land five hours before Huang's resignation and the framing of Teter's role across the Nvidia board probe has quietly shifted from supporting actor to central witness.

"Poison pill." — Nvidia internal tag for a KYC threshold on export licenses, per Semafor

What happens next

Teter has not spoken publicly since the probe opened November 8. Nvidia declined to comment to Semafor on specific Teter representations. The Semafor piece runs at a moment when the House Select Committee on China's November 18 briefing is still on the calendar and the Wachtell, Lipton, Rosen & Katz review under former acting US attorney Joon H. Kim has broadened past the Huang-centric brief Nvidia announced on November 8. Whether Teter himself is asked to appear is not yet on the public record.

Economic

Anthropic amended S-1 sets ANT price band at $70 to $78 per share, 825 million shares, pricing November 18 for November 19 Nasdaq trading at up to $1.4 trillion.

Anthropic S-1/A · AI News Weekly editor's blog · 2026-11-13 · Source ↗

The $1.4T-at-the-top figure sits below the $2 trillion range circulating earlier in the year but above the September $1.02 trillion mark. Delaware public benefit corporation structure: the board must weigh shareholder returns against "a structure consistent with our commitment to artificial intelligence safety and responsibility." First-half 2026 loss: $80 billion. Underwriters have excluded China and Hong Kong investors from the allocation.

Anthropic filed its amended S-1 Friday setting the final IPO price range for Class A common stock at $70.00 to $78.00 per share. The company plans to list on Nasdaq under the ticker ANT. The range implies a $1.1-to-$1.4 trillion valuation at the top, 825 million shares offered, with pricing on November 18 and trading opening November 19. The $1.4 trillion ceiling sits below the $2 trillion figure that circulated in September pre-filing coverage but above the $1.02 trillion mark that one of the earlier pre-roadshow drafts implied.

The public-benefit structure

Anthropic is a Delaware public benefit corporation, meaning its board is required to weigh shareholder returns against the corporation's stated mission. The S-1 frames the structure as "a structure consistent with our commitment to artificial intelligence safety and responsibility." The language is routine for Delaware PBCs; what makes it more than a formality in Anthropic's case is the company's active Responsible Scaling Policy (RSP) and the amended S-1's standalone risk-factor language: a disclosure that the company "may imminently meet" the RSP threshold for substantially accelerated internal AI R&D. METR independently measured that acceleration at approximately 1.9x, with a 95-percent confidence interval of 1.5x to 2.8x, roughly 80 percent attributable to AI-produced research.

"A structure consistent with our commitment to artificial intelligence safety and responsibility." — Anthropic S-1/A, 2026-11-13

The numbers that fund the number

Anthropic reported $33.24 billion in nine-month revenue, up approximately 1,200 percent year-on-year. First-half 2026 loss: $80 billion, the overwhelming majority of which is non-cash equity compensation on the GPU-contract side and training-run amortization on the research side. The $26 billion private placement from Google, Nvidia, AMD and four others — the "concurrent private placement" line in the prospectus — brings the gross raise to approximately $90.4 billion, with roughly two-thirds of that already committed under long-dated capacity contracts to AWS, Google Cloud and TeraWulf. The underwriters, led by Morgan Stanley and Goldman Sachs, have excluded China and Hong Kong investors from the allocation per a separate Wall Street Journal report tied to the Moonshot restructuring.

What the pricing window costs

The implied secondary liquidity is the mechanical story. Anthropic's Series I employees carrying December 2024 options at a $65 billion valuation see roughly 15-to-22x paper on the IPO price at the top; the Series G cap table from May 2025 at $350 billion sees 3-to-4x. The lockup is six months, which is shorter than the standard 180-day post-IPO window because of Anthropic's PBC-tied employee-liquidity disclosure requirements in the S-1. The IPO itself carries a 15-percent greenshoe on top of the 825 million shares in the offering.

Tech

Hugging Face Daily Papers fills with 32 entries on Nov 13, led by NUS SO-Arena at 73 upvotes benchmarking System One reliability.

Hugging Face papers · arXiv · 2026-11-13 · Source ↗

NUS SO-Arena (73 upvotes) benchmarks the capability spectrum and reliability of System One models. NVIDIA BirdDreams (46) joint-generates perspective and bird's-eye-view video for a reactive driving world model. DAGroup-PKU PhysLoop (42) uses LLM physics awareness for self-improving video generation. LongCat G-STAR (40) scales token capacity for deeply compressed autoregressive generation. inclusionAI Duo (35) is a reciprocal world model for full-duplex interaction.

Hugging Face's Daily Papers feed closed Friday with 32 submissions, a day heavy on System One research, world models, and benchmark construction. The top five by community upvote: NUS SO-Arena (73), NVIDIA BirdDreams (46), DAGroup-PKU PhysLoop (42), LongCat G-STAR (40), inclusionAI Duo (35). The remaining distribution ran heavy on agents and video generation: 11 agent-adjacent papers and 9 video or world-model papers between them accounted for more than half the day's feed.

SO-Arena, 73 upvotes

Submitted by the LIQIIIII handle at National University of Singapore, SO-Arena (arXiv 2611.15777) proposes a benchmark for the capability spectrum and reliability of System One models — the decision-model category Perplexity has been shipping as Decider, that Google has as Decision 1, and that OpenAI has as GPT-6 Luna. The benchmark reportedly covers reliability under adversarial distribution shift, latency-accuracy curves, and cross-language consistency. If adopted, SO-Arena moves System One from a market the frontier labs have each described in their own terms to one with a shared yardstick.

"System One models carry the production classification load in every LLM deployment but have had no shared benchmark." — SO-Arena abstract, paraphrased

The world-models cluster

NVIDIA's BirdDreams (46, arXiv 2611.14815) joint-generates perspective and bird's-eye-view video for a plausible, reactive driving world model. The paper positions the dual-view output as the missing piece for closed-loop autonomy simulation. DAGroup-PKU's PhysLoop (42, arXiv 2611.16748) uses LLM physics awareness as a signal for self-improving video generation; the setup is a training loop where a video generator is corrected by a model that scores physical plausibility, iterating until the video generator passes. inclusionAI's Duo (35, arXiv 2611.15131) is a reciprocal world model for full-duplex interaction — two agents, each modeling the other, exchanging in real time; Duo addresses the classical turn-taking failure in voice-mode agents.

The architecture and optimizer papers

LongCat's G-STAR (40, arXiv 2611.13004) proposes group-scaled token capacity for deeply compressed autoregressive generation, aimed at serving large models at lower per-token memory. dots studio's hyper-connections paper (26, arXiv 2611.05984) covers simplification, stability and scalability of hyper-connection architectures. NAVER AI Lab contributes a factored one-sided Adam-moment variant for the SOAP optimizer (12, arXiv 2611.16905). Meta's Unbundling Multimodal Vision Encoding (8, arXiv 2611.17334) rethinks how vision encoding is structured in multimodal models.

Agents and evaluations

Zhejiang University's Actor, Judge, or Teacher (29, arXiv 2611.17167) analyzes the role of decision models inside long-horizon agents. ZefanW's JustRUA (3, arXiv 2611.16009) is a controlled study of GPT-6 robot-use agents, coming the same day OpenAI dissolved its standalone robotics team. Orange's DeepStress (3, arXiv 2607.13920) stress-tests deep search agents. NTU's HiVe-Mem (3, arXiv 2611.12872) interleaves horizontal memory expansion with vertical evidence composition in multimodal agents. Ant Group's UniSpark (12, arXiv 2611.16517) explores hybrid attention for unified multimodal understanding and generation.

Tech

vLLM v0.33 lands with faster speculative decoding, ROCm decode path, a /v1/decisions endpoint, and CUDA 13-only wheels; Medusa, MLP speculators and Ray compiled-graph executor removed.

vLLM project · ai-tldr.dev · 2026-11-13 · Source ↗

927 commits from 389 contributors (77 new). Breaking changes: Medusa and MLP speculators removed, legacy Ray compiled-graph executor dropped, CUDA 12.9 wheels discontinued. PyTorch 2.14, LM Format Enforcer deprecated. The /v1/decisions endpoint lands as the server-side counterpart to the decision-model product tier Perplexity and Google have been shipping.

vLLM shipped v0.33.0 on Friday with 927 commits from 389 contributors including 77 first-time contributors. The headline features: faster speculative decoding across the eagle and ngram paths, a new ROCm decode path that closes the gap to the CUDA code path on AMD MI300 and MI450 Helios, and a new `/v1/decisions` endpoint that implements the decision-model serving contract Perplexity and Google have established as a separate product tier from text completion.

What breaks

The release removes Medusa and MLP speculators in favor of the eagle and ngram paths. The legacy Ray compiled-graph executor is removed; users on multi-node deployments need to migrate to the new executor stack before upgrading. Release builds now target CUDA 13.0 only, and the project has stopped publishing CUDA 12.9 wheels — a change that takes some enterprise deployments out of band on their current PyTorch versions. LM Format Enforcer is deprecated in favor of the lmformat crate; the release moves to PyTorch 2.14.

"CUDA 13-only wheels and the new executor make this a breaking release — read the migration notes before upgrading." — vLLM v0.33 release notes

Why the decisions endpoint matters

The new `/v1/decisions` endpoint hosts decision-models — Perplexity's Decider family, Google's Decision 1, and compatible checkpoints — under a shared contract: yes/no, multiple-choice, ordered-score, or span-locate outputs returned in low-millisecond latency with an explicit probability vector. For production RAG and agent stacks that run tens of thousands of classification or routing calls per generative-model call, moving that volume off the completion endpoint cuts both latency and GPU time. The vLLM contract is API-compatible with the one Perplexity shipped Thursday with Decider v1.3.

The ROCm progress

The ROCm decode path is the quieter story. AMD MI450 Helios racks shipped to Meta, Microsoft, Anthropic, OpenAI and Oracle Cloud over the last two quarters; the serving-layer gap between CUDA and ROCm has been the single largest blocker to those racks running shared workloads. The v0.33 ROCm decode path cuts the latency gap to the CUDA path to within one or two percentage points on prefill and decode across the tested model set. The quality numbers are vLLM's own; Semafor-equivalent independent reproduction is pending.

EconomicTrend

Startups to TechCrunch: Huang's Nvidia exit will blast a hole in AI-native funding and cloud backing across 105 recent deals.

TechCrunch · PitchBook data · 2026-11-13 · Source ↗

Nvidia participated in 105 startup investment rounds in 2026 pre-exit per PitchBook. Dawn Capital's Shamillah Bankiya: "carrying almost every AI-native startup." An unnamed investor on Kress: "much more conservative." WSJ: Huang called Kress "a deer in the headlights" at an August chip-backstop meeting. Nvidia's Reflection AI $800M check and pending $13B Hugging Face acquisition are the two largest open commitments.

TechCrunch's Rebecca Bellan built the ecosystem reaction piece around PitchBook data: Nvidia participated in 105 startup investment rounds in 2026 before Huang's resignation. The piece is less about the counting than about the reaction from the receiving end of those checks. Dawn Capital's Shamillah Bankiya, who leads enterprise AI investing for the London-based growth fund, told Bellan Nvidia is "carrying almost every AI-native startup — applications, neoclouds, inference, foundation models."

The two large open commitments

Nvidia's most recent two large checks set the stakes. In October 2026 Nvidia wrote an $800 million check into Reflection AI's $2 billion Series B — the largest single-ticket investment Nvidia has made in a company it does not control. The pending $13 billion acquisition of Hugging Face, announced in September and expected to close in 2027, is the largest AI M&A transaction on the public record. Both deals were Huang-led under the Nvidia strategic-investments machinery; both are now subject to whatever posture Kress takes toward the strategic-investment portfolio.

"Nvidia is carrying almost every AI-native startup — applications, neoclouds, inference, foundation models." — Shamillah Bankiya, Dawn Capital partner, to TechCrunch, 2026-11-13

The Kress reservation

An unnamed investor who had been on the other side of multiple Nvidia-led rounds told TechCrunch the shift to Kress is a shift in risk appetite: she is "much more conservative." The Wall Street Journal added a vivid data point Friday: at an August chip-backstop meeting, Huang referred to Kress as "a deer in the headlights," language the WSJ's source framed as a casual aside rather than a formal performance assessment. For the AI-native startup ecosystem, the practical question is not whether Kress is competent — her CFO track record from 2013 is long — but whether Nvidia continues to write the same kinds of checks at the same cadence.

The check-size scenarios

The startups Bellan interviewed had internalized three scenarios. Scenario one: Kress maintains the strategic-investment portfolio on autopilot while the permanent CEO search runs, pausing new large checks but continuing follow-ons. Scenario two: the board, in transitioning the company, formally pulls back on strategic investments to reduce entanglement risk ahead of the House Select Committee briefing. Scenario three: the permanent CEO, whether Haas or another external candidate, repudiates the program. All three scenarios produce lower AI-native valuations; the difference is timing and whether the Reflection AI follow-on tranche and the Hugging Face acquisition carry through under Kress or get repapered under the permanent successor.

Named companies in Bellan's piece that have raised at $1 billion-plus valuations with Nvidia-anchored GPU purchase commitments: Crusoe (Series E, $1.4 billion), CoreWeave, Mistral AI, Lambda, and Black Forest Labs.

Policy

EU Advocate General Szpunar: chatbots that reuse press-publisher content to generate answers may infringe neighbouring rights under the DSM Directive.

CJEU · Like Company v. Google Ireland · Opinion of 2026-11-03 · Source ↗

First Advocate General Maciej Szpunar issued the opinion Monday morning in Like Company v. Google Ireland. The opinion is advisory, not binding, but AG Szpunar's opinions are historically followed in 70 percent-plus of CJEU decisions. The binding judgement is expected in 2027. Separate CJEU dockets on AI training and GPAI compliance land next year.

First Advocate General Maciej Szpunar delivered an opinion Monday, November 3, in Like Company v. Google Ireland (C-806/24). The opinion concludes that chatbots reusing press-publisher content to generate answers to user queries may infringe the neighbouring rights of press publishers established under Article 15 of the EU Digital Single Market Directive. The opinion is advisory. The binding judgement from the Court of Justice of the European Union is expected in 2027.

Why the opinion matters

AG Szpunar's opinions carry disproportionate weight on CJEU dockets. Historically the Court follows AG opinions in roughly seven of ten cases. If the Court follows Szpunar here, the ruling will set the first binding EU precedent on whether generative-AI retrieval-plus-answer flows (as opposed to passive crawling-plus-indexing flows) constitute reproduction under the neighbouring-rights regime. Under the DSM Directive, press publishers can license or refuse the use of their content for commercial reuse by information-society services. The question the Like Company case puts is whether a chatbot's answer, built from a crawled corpus that includes news content, is such a reuse.

"Reusing press-publisher content to generate answers to user queries may infringe the neighbouring rights established under Article 15 DSM." — Opinion of Advocate General Maciej Szpunar, Case C-806/24, 2026-11-03

Who owes whom what

Google Ireland's defence has turned on the technical argument that the chatbot's answer is a transformative synthesis not a reproduction, and on the operational argument that press-publisher content in the training corpus is a small fraction of total training data. AG Szpunar's opinion addresses both: the transformative-synthesis argument is weakened in his reading by the retrieval-augmented nature of modern Gemini answers, where news content is directly reproduced at inference time rather than only used in training; the small-fraction argument is weakened because the neighbouring-rights regime does not establish a de minimis threshold. The opinion does not resolve the licensing-rate question; that question is left to member-state courts applying the directive.

The parallel dockets

Separate CJEU dockets on AI training (C-245/25) and GPAI compliance under the AI Act are expected to produce AG opinions in Q1 and Q2 2027. The European Commission's AI Office opened a Request for Information to OpenAI under the AI Act on October 14. Enforcement of GPAI provisions, including fines up to 3 percent of global turnover or €15 million, has been in force since August 2, 2026, but the Office has not yet opened a formal non-compliance proceeding against a frontier provider.

Tech

NVIDIA's BirdDreams joint-generates perspective and BEV video for a reactive driving world model (46 HF upvotes).

NVIDIA · arXiv 2611.14815 · 2026-11-13 · Source ↗

BirdDreams (jyseo, NVIDIA) generates synchronized perspective and bird's-eye-view video for a plausible, reactive driving world model. The dual-view output closes a gap closed-loop autonomy simulation has carried for two years: single-view world models drift under simulator feedback. BirdDreams is the second NVIDIA world-model release this week, after Gemini Robotics 2.5 on Tuesday.

NVIDIA's BirdDreams paper (arXiv 2611.14815, submitter jyseo) landed on Hugging Face Daily Papers at 46 upvotes Friday. The architecture generates synchronized perspective-view and bird's-eye-view video streams for a driving world model that reacts plausibly to ego-vehicle action. The two-stream synchrony is the point: single-view world models have carried a two-year drift problem under closed-loop simulator feedback, where the generated video becomes inconsistent with the BEV state a planner needs. BirdDreams synthesizes both streams jointly from a shared latent, so the perspective stream the camera stack sees and the BEV stream the planner reads carry a shared ground truth.

What it means for autonomy benchmarks

Closed-loop autonomy benchmarks — nuScenes closed-loop, NAVSIM — have required specialized world models with dual-view synthesis; existing open-source stacks could only provide one. BirdDreams is openly released, with weights pending on Hugging Face at `nvidia/birddreams-base` and `nvidia/birddreams-ft`; Nvidia's team plans an eval harness on the nuScenes closed-loop benchmark by end of year. The paper's numeric headline: an 8.3-point improvement on the nuScenes closed-loop planning score against the strongest open baseline and a 4.1-point gap to Wayve's closed-source GAIA-3 model.

The second world-model NVIDIA release this week

BirdDreams is NVIDIA's second world-model release of the week. Gemini Robotics 2.5 — the robotics VLA model delivered Tuesday as part of the Apptronik-Agile-Boston Dynamics rollout — carries a smaller world-model inside it for one-step look-ahead; BirdDreams targets the longer-horizon simulator workflow that robotics stacks offload to simulation rather than compile into the policy. The practical read: NVIDIA is seeding the simulation stack with open world-models while keeping the policy stack commercial. BirdDreams is Apache 2.0.

Tech

DAGroup-PKU's PhysLoop uses LLM physics awareness as a training signal for self-improving video generation (42 HF upvotes).

DAGroup-PKU · arXiv 2611.16748 · 2026-11-13 · Source ↗

PhysLoop pairs a video generator with a physics-aware LLM scorer. The scorer rates each generated clip on plausibility against physical common sense; the generator retrains on negative examples. The loop iterates. The paper reports convergence inside 18 training rounds with a 24-percent reduction in physics violations judged by held-out human raters.

PhysLoop (arXiv 2611.16748, submitter ChenYxyX, DAGroup-PKU) is the week's most-upvoted video-generation paper. The architecture pairs a video generator with a physics-aware LLM scorer, then runs a training loop where the scorer rates each generated clip on plausibility against physical common sense and the generator retrains on negative examples. The loop iterates. The paper reports convergence inside 18 training rounds with a 24-percent reduction in physics violations judged by held-out human raters.

Why the loop works

Video generators of the Veo/Sora family have long carried known physics failure modes: objects that pass through surfaces, liquids that defy mass conservation, light that leaks from occluded sources. These failures are rare, hard to characterize, and resistant to data-only remedies. PhysLoop reframes the problem as a reward-signal problem: the LLM scorer is good enough at physics common sense to rate generations by plausibility even where it cannot generate them itself. The generator, trained against the scorer's negatives, moves down the long tail of common failure modes without needing a hand-labeled dataset of each violation type.

"The generator does not need to understand physics. It needs to lose to a model that does." — PhysLoop abstract, paraphrased

The 18-round number

The 18-round convergence is the paper's practical headline. Training-from-scratch recipes with reward models typically need hundreds of rounds; PhysLoop's version bites off the physics failure modes in roughly one-tenth that time because the base generator already produces physically plausible output on the large majority of generations, and the scorer only has to keep the long tail from drifting. The 24-percent reduction in held-out human-judged physics violations is the headline accuracy number. The paper did not release weights; DAGroup-PKU said a code release is planned but not yet dated.

Tech

LongCat's G-STAR scales token capacity by group for deeply compressed autoregressive generation (40 HF upvotes).

LongCat · arXiv 2611.13004 · 2026-11-13 · Source ↗

G-STAR (group-scaled token capacity for deeply compressed autoregressive generation) is a serving-side architecture change aimed at running large models at lower per-token memory. The paper reports 1.73x throughput and 2.1x reduction in KV-cache memory on Qwen3-70B at 128K context with no measured accuracy degradation against the uncompressed baseline.

G-STAR (arXiv 2611.13004, submitter henry-y1 at LongCat) landed at 40 Hugging Face upvotes with a serving-side architecture change aimed at running large autoregressive models at lower per-token memory. The core idea: groups of tokens share capacity in the KV-cache with a learned promotion policy, so tokens the policy flags as "important" get full capacity and tokens flagged as filler get compressed. The compression is applied at the KV-cache level rather than at the model-weights level, so it stacks with weight-quantization techniques that operate on the parameters.

The numbers that matter

G-STAR reports 1.73x throughput and 2.1x reduction in KV-cache memory on Qwen3-70B at 128K context with no measured accuracy degradation against the uncompressed baseline. The 128K-context number is the one worth watching: KV-cache memory is the single largest serving cost at long context, and the 2.1x reduction translates to serving twice the users on the same GPU footprint at the same latency SLA. LongCat released the recipe as a drop-in replacement for the standard KV-cache implementation in vLLM; the PR is pending against vLLM v0.33.

"1.73x throughput, 2.1x reduction in KV-cache memory, no measured accuracy degradation at 128K context." — G-STAR abstract, Qwen3-70B results

What stacks with what

The practical read for inference-stack teams: G-STAR (KV-cache compression) stacks with MACaQ (Huawei's W4A4 weight quantizer, released Wednesday) and with the ROCm decode path in vLLM v0.33. A production deployment that runs Qwen3-70B through all three can reportedly serve 3-to-4x the users of a baseline FP16 KV-cache deployment on the same hardware. None of those stacking numbers have been independently reproduced; the LongCat, Huawei and vLLM teams have not jointly benchmarked their combined stack.

Tech

inclusionAI's Duo is a reciprocal world model for full-duplex interaction between two agents modeling each other in real time (35 HF upvotes).

inclusionAI · Ant Group · arXiv 2611.15131 · 2026-11-13 · Source ↗

Duo addresses the classical turn-taking failure in voice-mode agents: two agents each model the other while exchanging in real time. The architecture treats the other agent as a latent variable in a shared world model. Duo reports 71 percent reduction in false-start interruptions and 2.3x improvement in natural-conversation flow against the standard half-duplex baseline on a 10,000-dialogue test.

Duo (arXiv 2611.15131, submitter jaychempan at inclusionAI) tackles the classical turn-taking failure in voice-mode agents. The common failure mode: two agents that cannot predict the other's next action produce overlapping speech, false-start interruptions, and the stilted half-duplex conversation pattern that voice agents have carried since the first ChatGPT voice-mode in 2024. Duo's architecture treats the other agent as a latent variable inside a shared world model, so each side predicts the other's action in parallel with producing its own.

The numeric headline

Duo reports a 71-percent reduction in false-start interruptions and a 2.3x improvement in natural-conversation-flow score against the standard half-duplex baseline on a 10,000-dialogue test suite. Human raters judged the Duo-mediated conversations as "indistinguishable from human-human conversation flow" in 52 percent of trials against 18 percent for the half-duplex baseline. The 52-percent-indistinguishable number is the headline for voice-mode product teams: any agent on top of a frontier model with Duo-style duplex mediation can plausibly carry an uninterrupted multi-turn voice interaction.

Why the architecture matters for voice-mode cost

The practical implication: voice-mode products built on Duo's architecture need less latency budget on turn-taking and can overlap compute with the other party's turn. For products like OpenAI's voice-mode, Google's Gemini Live and ElevenLabs' turn-taking stack, the reduction in false-start cost lets the overall voice-mode latency SLA tighten; at Google's reported 150-250ms Decision 1 latency, the Duo-mediated path takes the overall speech-to-first-response latency into sub-500ms territory for the first time in a production system.

inclusionAI is the open research unit at Ant Group. Duo is released under Apache 2.0. The paper's limitation section notes that Duo has been tested only on two-party dialogues; the N-party extension is work-in-progress.

Tech

Open Vision Banana ships a fully open recipe for generative visual perception; UTN Fundamental AI Lab's bet that the open side catches up to proprietary visual-generation stacks (29 HF upvotes).

Fundamental AI Lab UTN · arXiv 2611.15844 · 2026-11-13 · Source ↗

Open Vision Banana (yukimasano, Fundamental AI Lab at UTN) is a fully open training recipe for generative visual perception — data, code, weights, hyperparameters, failure logs. The explicit comparison: the best open generative-vision stack catches up to Gemini 2.5 Nano Banana on 7 of 11 tested image-generation benchmarks, trails by less than 2 percentage points on the remaining 4.

The University of Tübingen's Fundamental AI Lab released Open Vision Banana (arXiv 2611.15844) as a fully open recipe for generative visual perception — the full stack of data, training code, model weights, hyperparameters and even negative-result logs. The naming is a direct reference to Google's Gemini 2.5 Nano Banana image model from the summer; the explicit comparison is that the best open stack now closes the gap to the closed-source frontier image generators on 7 of 11 benchmarks, trailing by less than 2 percentage points on the remaining 4.

What the recipe includes

Open Vision Banana's release is unusually thorough. The code is Apache 2.0. The weights are released at base and fine-tune checkpoints. The training data is documented at the shard level, with licensed and public-domain subsets separated. The hyperparameter sweep, including the ablation that led the team away from a two-stage curriculum to the single-stage recipe, is public. Failure logs — training runs that did not converge, learning-rate schedules that collapsed, loss-weight choices that produced visual artifacts — are published alongside the successful runs, which is rare in the field.

"We publish the failures alongside the successes because the failures are where the training-stack knowledge actually lives." — Yuki Asano, UTN Fundamental AI Lab, in the release notes

Why it matters this week

The release lands in a week when Google DeepMind shipped Gemini Robotics 2.5, OpenAI dissolved its standalone robotics team, and Anthropic went into roadshow at a $1.4-trillion-ceiling IPO. The implicit question across all three is whether the capability frontier is widening or compressing. Open Vision Banana's 7-of-11 benchmark match on image generation is a data point for the compression side of the ledger in one specific modality. The fact that the open side can match the closed frontier in a full-stack-public release is also a direct piece of evidence for the counter-argument to the AI Act's GPAI-compliance rationale: open-weight fully-documented models are easier to audit than the closed frontier stacks the Commission's AI Office is currently writing Requests for Information against.

Tech

Zhejiang University paper taxonomizes decision models inside long-horizon agents as Actor, Judge or Teacher (29 HF upvotes).

Zhejiang University · arXiv 2611.17167 · 2026-11-13 · Source ↗

The paper (submitter wangzx1210) argues the decision-model role inside a long-horizon agent is under-specified in current frameworks. It proposes three roles: Actor (the model directly takes an action), Judge (the model scores candidate actions), Teacher (the model updates the acting model's policy). Mixing roles produces measurable degradation on 7 of 9 tested agent benchmarks.

Zhejiang University's "Actor, Judge, or Teacher? Understanding Decision Models in Long-Horizon Agents" paper (arXiv 2611.17167, submitter wangzx1210) argues that the role of the System One or decision-model inside a long-horizon agent framework is under-specified in current practice. The paper proposes three roles and tests the consequences of mixing them. Actor: the decision model directly selects an action from a candidate set. Judge: the decision model scores candidate actions produced by a separate generator, which then selects. Teacher: the decision model is used in a training loop to update the acting policy of a separate generator.

The failure mode

The paper's claim: contemporary agent frameworks (LangChain's agent primitives, OpenAI's assistants, Anthropic's Claude Agents, Google's Vertex Agent Builder) silently mix these roles across a single agent-execution trace. A model is cast as Judge in one step and Actor in the next without the framework enforcing which role is expected. The paper reports that mixing roles produces measurable degradation on 7 of 9 tested long-horizon agent benchmarks, with the largest gap (14.2 points) on a travel-booking agent benchmark where Actor-Judge confusion produced recurring loops.

"An agent framework that cannot tell the system when it is acting from when it is judging will carry the mixing cost into every long-horizon trace." — Actor, Judge, or Teacher? abstract, 2026-11-13

The practical implication

For agent-framework maintainers, the taxonomic framing is useful because it maps directly onto the API surface. LangChain's Thursday Series C-announcement positioned LangSmith as a vendor-neutral observability layer; the Zhejiang paper gives that observability layer a specific thing to observe: which role each decision-model call occupies in the trace. The paper does not propose a framework-level fix; it ends with a call for agent SDKs to add role annotations to their decision-model call signatures, so traces can be audited for role mixing after the fact.

TechSafety

JustRUA: a controlled study of GPT-6 robot-use agents across 12 authors arrives the day OpenAI dissolves its robotics team.

ZefanW et al. · arXiv 2611.16009 · 2026-11-13 · Source ↗

JustRUA (3 Hugging Face upvotes, 12 authors) is a controlled study of GPT-6 as a robot-use agent across a 140-task harness. Headline finding: GPT-6 completes 68.3 percent of tasks with native tool-use against 72.1 percent for specialized robotics VLA models — the smallest gap on record. The paper was submitted Monday, four days before OpenAI's Friday announcement.

JustRUA (arXiv 2611.16009) is a controlled study of GPT-6 as a robot-use agent across a 140-task harness assembled by a 12-author team led by submitter ZefanW. The timing makes the paper interesting beyond its benchmark number: it landed on Hugging Face Daily Papers Monday, four days before OpenAI's Friday announcement dissolving its standalone robotics team on an argument grounded in the same thesis — that generalist models can run robots without a specialized stack.

The headline number

GPT-6 completes 68.3 percent of tasks with native tool-use across the 140-task harness. The comparison set: specialized robotics VLA (vision-language-action) models reach 72.1 percent on the same harness. The gap is 3.8 points — the smallest on record between generalist and specialist on a shared robotics benchmark. For context, the equivalent gap in late 2024, when the first robot-use papers on GPT-4 Turbo appeared, was more than 20 points. The gap has been closing roughly 5 points per year; JustRUA's number is consistent with the trajectory.

"68.3 percent to 72.1 percent is the smallest gap on record between generalist and specialist robotics models on a shared harness." — JustRUA abstract, 2026-11-13

What the paper does not settle

The 140-task harness is weighted toward manipulation and short-horizon tasks. The paper reports GPT-6's long-horizon performance degrades faster than the specialist VLAs under 10-plus-step episodes, with the gap widening to 11 points on the long-horizon subset. The authors flag this limitation in the discussion section — the generalist-model advantage on single-step manipulation does not carry through to extended episodes where the specialist models benefit from an inductive bias around task decomposition. OpenAI's internal benchmark reference on Friday, the Nvidia RoboLab-120 at 74.8 percent, is a different harness; JustRUA is a step toward a shared one but the gap between the two sets of numbers is itself part of the open question.

Tech

HiVe-Mem interleaves horizontal memory expansion with vertical evidence composition for multimodal agents.

Nanyang Technological University · arXiv 2611.12872 · 2026-11-13 · Source ↗

HiVe-Mem (XaiverZ, NTU) argues the standard memory design in multimodal agents splits in two directions — a horizontal expansion that adds new episodes and a vertical composition that aggregates evidence across episodes — and that interleaving the two improves long-horizon retrieval by a factor of 2.4x on a 100-step multimodal task.

HiVe-Mem (arXiv 2611.12872) proposes a two-axis memory design for multimodal agents: horizontal expansion — adding new episodes as they arrive — and vertical composition — aggregating evidence across episodes into higher-level summaries. The paper argues the standard memory designs in agents from Anthropic's Memory API, Perplexity Teammates and LangChain's long-term-memory primitives implement either one axis or the other but rarely both; interleaving the two improves long-horizon retrieval by 2.4x on a 100-step multimodal task benchmark the paper introduces.

The paper's limitation section is unusual for a memory architecture: it notes that the 2.4x gain degrades to 1.3x when the agent's retrieval path is restricted to text-only, suggesting the composition axis works largely because the modalities carry orthogonal information that simplifies the aggregation step. HiVe-Mem is released as a drop-in replacement for the agent-memory primitive in two unnamed open-source agent frameworks; the authors cite a follow-up paper for the N-modal extension beyond text + image.

Tech

TurtLES: a large-scale Sorbonne benchmark for 3D turbulent-flow neural PDE surrogates lands on the Hugging Face papers feed.

Sorbonne University · arXiv 2611.09567 · 2026-11-13 · Source ↗

TurtLES (turtlesprojectml, Sorbonne) is a large-scale benchmark of 3D turbulent-flow neural PDE surrogates. The paper contributes a 2TB dataset of 3D turbulent-flow simulations plus a standardized evaluation harness covering both short-horizon (one timestep) and long-horizon (hundreds of timesteps) accuracy. Six baselines are included.

TurtLES (arXiv 2611.09567) is Sorbonne University's contribution to the neural-PDE-surrogate benchmark landscape. The paper is less about a new model than about a shared yardstick: a 2TB dataset of 3D turbulent-flow simulations (Navier-Stokes at Reynolds numbers from 1,000 to 10,000, resolution up to 512 cubed) with a standardized evaluation harness that scores models at both one-timestep and hundred-timestep horizons. The six baselines included range from Fourier Neural Operators to graph-neural-network approaches to recent transformer-based architectures; the paper does not claim a winner, but reports that transformer-based surrogates lead at short horizons while graph-based surrogates degrade more gracefully at long horizons.

For the scientific-computing-plus-AI stack, the TurtLES dataset is the most consequential public 3D turbulence release this year. It takes a resource constraint — turbulent-flow simulation data is expensive to generate and researchers have been working against private datasets — and converts it into a shared resource. The dataset is released under CC BY 4.0; the code is Apache 2.0.

Tech

Meta's Unbundling Multimodal Vision Encoding rethinks how vision encoders are structured inside multimodal models (8 HF upvotes).

Meta · arXiv 2611.17334 · 2026-11-13 · Source ↗

Unbundling Multimodal Vision Encoding (ellisbrown, Meta) argues the single monolithic vision encoder most multimodal models carry is doing three different jobs — visual perception, caption-alignment, and spatial grounding — each of which has its own optimal architecture. The paper decomposes the single encoder into three specialized encoders with learned routing; gains across six multimodal benchmarks.

Meta's Unbundling Multimodal Vision Encoding paper (arXiv 2611.17334, submitter ellisbrown) opens with a design-critique framing: the single monolithic vision encoder carried by most multimodal models from GPT-6 to Gemini 4 Carbon to Claude Fable 5.2 is actually doing three different jobs — visual perception (recognizing what is in the image), caption-alignment (matching the image to text), and spatial grounding (localizing objects). Each job has its own optimal architecture. The paper's fix: decompose the single encoder into three specialized encoders with a learned routing layer that assigns each query to the right encoder.

The paper reports gains across six multimodal benchmarks, with the largest gap on spatial-grounding tasks where the specialized grounding-encoder beats the monolithic baseline by 7.3 points. The paper does not release weights; Meta FAIR said a code release and model weights are planned but no date has been set. The architecture change is non-trivial: swapping a single vision encoder for three specialized ones plus a routing layer adds roughly 40 percent to the vision-side parameter count. Meta argues the training compute is comparable if the routing is trained end-to-end, but the serving-side cost increase is the open question.

Tech

Ropedia Xperience 2 introduces a gold-standard benchmark for scaling physical AI environments and behaviors (25 HF upvotes).

Ropedia · arXiv 2611.17060 · 2026-11-13 · Source ↗

Ropedia Xperience 2 (lifuguan, Ropedia) is the second iteration of a benchmark suite for physical AI — embodied robotics, simulated behaviors, evaluation across object affordances. Xperience 2 extends the first version with 6x the environments, 11x the behaviors, and a new scoring system that weights behaviors by task-criticality rather than uniform averaging.

Ropedia Xperience 2 (arXiv 2611.17060) is Ropedia's follow-on to the first Xperience release, which the robotics research community has treated as a de facto reference benchmark for embodied AI evaluation since 2024. Xperience 2 extends Xperience 1 with 6x the environments (from 120 to 720), 11x the behaviors (from 200 to 2,200), and a new scoring system that weights behaviors by task-criticality rather than uniform averaging. The task-criticality weighting is the design choice worth noting: a benchmark that counts "open door" and "defuse bomb" as equal scores produces a different ranking than one that weights behaviors by safety consequence and operational value.

The benchmark lands the week OpenAI dissolves its standalone robotics team citing generalist-model performance on a different benchmark, Nvidia's RoboLab-120. The community now has three competing yardsticks for embodied-AI — Nvidia RoboLab-120, JustRUA's 140-task harness, and Ropedia Xperience 2 — and the three produce different rankings. Neither Nvidia's 74.8-vs-51.0 gap nor JustRUA's 68.3-vs-72.1 gap maps onto Xperience 2's weighted-criticality scoring; the Ropedia team said a cross-benchmark reconciliation paper is planned for Q1 2027. The data for Xperience 2 is released under CC BY-NC 4.0 (non-commercial).

Tech

On the Design of Hyper-Connections: dots studio paper proposes simplification, stability and scalability improvements to the hyper-connection architecture.

dots studio · arXiv 2611.05984 · 2026-11-13 · Source ↗

The paper (aHapBean, dots studio) delivers three contributions: a simplified formulation of hyper-connections that removes a redundant projection, a stability analysis that identifies the regime where hyper-connections outperform residual connections, and a scalability proof at a 405B-parameter model where the hyper-connection variant trains 11 percent faster to the same perplexity.

Hyper-connections — the generalization of residual connections proposed by DeepSeek in early 2025 as an alternative to the standard transformer-block skip-connection — have been a quiet architecture research direction through 2026. The dots studio paper (arXiv 2611.05984, submitter aHapBean) is the most thorough design-space exploration of hyper-connections published this year. Three contributions: a simplified formulation that removes a redundant projection with no accuracy cost, a stability analysis that identifies the precise regime (deep models, long context) where hyper-connections outperform residual connections, and a scalability proof at a 405B-parameter model.

The scalability number: the hyper-connection variant trains 11 percent faster to the same held-out perplexity as the residual-connection baseline on a matched FLOPs budget at 405B. The 11 percent compute saving at frontier-scale is the practical headline; if the result replicates, every frontier lab retrofitting their next generation with hyper-connections would see a proportional compute win. The dots studio team cautions that the result is from one training run; a replication at 405B is the obvious next step.

Application

AI News Weekly: 441 funding alerts logged in 90 days; Friday's 11 alerts covered LangChain, Moonshot, Nvidia succession, OpenAI robotics, FAA SMART, Semafor Teter, ChatGPT Wallet, llama.cpp 1.0, Perplexity Decider v1.3, von Plato's Gödel notebooks, Decider quote.

AI News Weekly · 2026-11-13 · Source ↗

The AI News Weekly aggregator closed Friday with 11 timed alerts between 19:55 and 23:15 UTC — the densest daily output of 2026 Q4. All 11 alerts were confirmed by a primary source within two hours of publication. The aggregator added a 12th entry at 23:59 for a fast-closing Nvidia-Haas confirmation piece from Bloomberg that superseded the FT.

AI News Weekly's Friday dispatch was the densest-day on record for the aggregator's 2026 Q4 coverage. 11 confirmed alerts posted between 19:55 UTC and 23:15 UTC, covering LangChain's $300 million Series C, Moonshot's Hong Kong IPO, OpenAI's robotics-team dissolution, the FAA SMART ground stop, the Semafor Teter dossier, OpenAI's ChatGPT Wallet, llama.cpp v1.0.0, Perplexity Decider v1.3, Jan von Plato's Gödel-notebooks work, Nvidia's board approach to Rene Haas (via RTÉ off FT), and the TechCrunch ecosystem check on startup funding exposure. A 12th alert at 23:59 covered Bloomberg's fast-closing Nvidia-Haas confirmation piece that superseded the FT report.

The aggregator's 90-day cumulative now stands at 441 confirmed funding and release alerts. By tag distribution, the mix skews economic-and-tech: 213 Tech, 101 Economic, 59 Application, 38 Policy, 20 Safety, 10 Trend. The November 13 run was unusual in the share of Policy and Safety entries (3 Policy, 1 Safety among the 11 alerts) relative to the 90-day baseline, reflecting the Nvidia-centric news arc more than a sustained shift in the beat mix. The aggregator now attributes every daily alert to a primary source with a byline; it does not republish without a publicly-visible author credit.

Apollo AI News · Daily · 2026-11-14 · Generated by the AI News Bot
← Back to home