Daily2026-09-16
Tech

Nvidia opens AI Infra Summit with 30x tokens-per-megawatt on Vera Rubin NVL72 and DSX MaxLPS.

Ian Buck's 9:55 a.m. keynote in Santa Clara Tuesday walked 8,000 attendees through the energy layer of the Vera Rubin platform. Nvidia said Vera Rubin NVL72 delivers up to 30 times higher throughput per megawatt than GB300 NVL72 on agentic workloads and enables up to 40 percent more GPUs inside the same power envelope. DSX MaxLPS, a new factory-wide scheduling layer, adds 1.4x more tokens per megawatt on top of that, and Lambda demoed a 24 percent cluster-wide token bump on stage using 19 nodes inside a 16-node budget. Amazon's Annapurna Labs signed on as the first NVHBM customer through Trainium4; d-Matrix put Raptor XPUs into NVLink Fusion with Vera CPUs; Emerald AI and Silicon Valley Power demonstrated automated AI-factory load-shed responding to grid signals in real time. Grace CPU already prints $5 billion in trailing revenue, and Nvidia is guiding data-center CPU revenue to more than double next fiscal year.

Tech · Nvidia Blog · 2026-09-15 · Source ↗

The AI Infra Summit opened Tuesday at the Santa Clara Convention Center with an attendee count that jumped from 3,500 in 2025 to more than 8,000 this year, making it the largest US-based AI-infrastructure conference of 2026 by a wide margin. Nvidia's Ian Buck, VP of hyperscale and HPC, drew the 9:55 a.m. main-stage slot and used it to move the pacing debate away from capability and toward power. Vera Rubin NVL72, now shipping into hyperscalers, delivers up to 30x higher throughput per megawatt than GB300 NVL72 on agentic workloads, up to 40 percent more GPUs inside an existing power budget, and up to 35 percent higher token throughput without new grid interconnects.

DSX MaxLPS is the layer on top. A factory-wide power-optimization scheduler that Nvidia said boosts performance per watt by a further 23 percent, bringing headline efficiency to roughly 1.4x more tokens per megawatt above the raw Vera Rubin gain. Lambda's live demo swapped the numbers into a concrete result: a 19-node cluster running inside a 16-node power envelope produced 5 million tokens per second, up from 4 million, a 24 percent lift with no capex.

"AI factories are becoming the new industrial platforms. Every watt now translates to a business outcome, and Vera Rubin is engineered to convert power into intelligence faster than anything before it." — Ian Buck, VP of hyperscale and HPC, Nvidia

Partner news around the keynote

Amazon's Annapurna Labs signed on as the first NVHBM customer through the Trainium4 line, giving AWS access to Nvidia's high-bandwidth-memory roadmap on a chip AWS designs itself. D-Matrix integrated its Raptor XPUs directly into NVLink Fusion with Vera CPUs, meaning an alternative training accelerator now shares Nvidia's coherent memory fabric rather than talking to Vera over PCIe. Emerald AI and Silicon Valley Power staged an on-stage demonstration of an automated AI-factory load-shed responding to hundreds of grid demand signals while holding critical workloads flat, the first public grid-integrated AI-factory demo built on production Nvidia stack. Perplexity reported 1.9x faster sandbox starts for secure agentic AI on Vera CPU, Redpanda cited 5.5x lower latencies and 73 percent higher throughput, Starburst claimed 3x faster query throughput, and Kinetica reported 2.7x on analytical queries.

Why it hit the pacing story

The Nvidia case against pacing is not a rebuttal. It is a reframing. Dario Amodei's essay treats the frontier as a capability curve to be paced. Buck argued from the stage that the tokens-per-megawatt curve is a separate axis where every hyperscaler already faces a hard power ceiling, and that gains on that axis compound whether or not capability racing slows. Grace CPU generated $5 billion in trailing-twelve-month revenue as of fiscal Q2 2027; Nvidia guided data-center CPU revenue to more than double this fiscal year to next; Vera adopters span Anthropic, OpenAI, xAI, ByteDance, CoreWeave and Oracle Cloud Infrastructure. Intel CEO Lip-Bu Tan followed Buck at 10:15 with a fireside titled "The AI Opportunity: Silicon to Systems," a defensive framing that named 18A foundry, Xeon 6, Gaudi 3 and rack-scale integration as the pieces Intel intends to hold together.

"Vera hardware delivers performance and efficiency that significantly exceed those of any other data center CPU in key areas." — Nvidia, quoted in Motley Fool coverage of the Vera CPU ramp

The tape did not buy the reframe cleanly. Nvidia gave back roughly 3.4 percent on Monday and closed near flat Tuesday even as AMD picked up 2 percent on Anthropic's confirmed MI450 deployment. The story markets are pricing is that the pacing coalition will slow demand faster than the tokens-per-megawatt curve can compensate. Buck's keynote is Nvidia's response.

Big Picture

Tuesday made the pace-the-frontier week concrete. Nvidia used AI Infra Summit to reframe the debate as tokens-per-megawatt rather than capability-per-quarter; Google DeepMind used the same day to ship Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking, taking first place on the Artificial Analysis Speech-to-Speech Quality Index at 82.6 ahead of GPT-Live 1 Astra and Grok-Voice Think Fast 2.0. Bloomberg confirmed that OpenAI, Anthropic and Google DeepMind have been coordinating directly on safety for weeks and do not believe an antitrust waiver is required, which pushed the pacing coalition from four CEOs on record to three labs in active operational contact. Shanghai AI Laboratory formally released Atria Dawn Preview, a 744B-parameter agentic mixture-of-experts on MIT-licensed weights, extending the open-weights frontier deeper into agentic territory just as the closed-source coalition tightened.

Anthropic's IPO advanced. Bloomberg confirmed the Nasdaq listing at up to $2 trillion, with Nvidia in talks to anchor at $10 billion and Morgan Stanley, Goldman Sachs and JPMorgan running the deal. Anthropic separately shipped Claude for Financial Advisors at the Future Proof Festival with BlackRock, Charles Schwab and Vanguard among 11 launch integrations, giving the roadshow a concrete vertical to point at. Salesforce opened Dreamforce with the Siemens deployment of Agentforce to 18,000 sellers in 132 countries; Sam Altman is scheduled with Marc Benioff at 2 p.m. PT for the first shared main stage since the pacing story broke. AIUC came out of stealth with $55 million from Ribbit, Nat Friedman and Anthropic co-founder Ben Mann to certify individual deployed agents against a SOC-2-style standard.

The counter-narrative did not soften. Trump called the CEO regulatory calls a "conspiracy"; Speaker Mike Johnson told the caucus the response is a White House convening, not legislation; John Kennedy is drafting a kill-switch amendment. Time named an OpenAI researcher, Marcus Williams, publicly estimating 70 percent extinction risk absent regulation, the highest on-record number from a currently-employed frontier-lab researcher. Bilal Chughtai made his July departure from Google DeepMind's AGI safety team public with an X post arguing frontier labs are racing "manic" toward a finish line society cannot absorb. Hugging Face's daily papers board surfaced two independent recursive-self-improvement papers in the top ten (Dream-RSI, RSIAgent) on the same day.

The rundown

Tech

Google ships Gemini 3.8 Live and takes speech-to-speech leaderboard at 82.6.

Officechai / Google DeepMind · 2026-09-15 · Source ↗

Google DeepMind rebuilt its live-audio stack around two models Tuesday. Gemini 3.8 Live Extended Thinking scored 82.6 on Artificial Analysis Speech-to-Speech Quality, ahead of GPT-Live 1 Astra and Grok-Voice Think Fast 2.0, and both variants detect and switch among 97 languages mid-conversation.

Google DeepMind released Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking on Tuesday, splitting its real-time voice stack into a cost-efficient conversational model and a higher-complexity variant that runs multi-step reasoning without interrupting dialogue. Both models switch among 97 languages mid-conversation, execute tool calls and API requests in the background while still speaking, and process visual input in near real time. Gemini 3.8 Live Extended Thinking took first place on the Artificial Analysis Speech-to-Speech Quality Index at 82.6, ahead of OpenAI's GPT-Live 1 Astra and xAI's Grok-Voice Think Fast 2.0. It also posted 68.6 percent on τ-Voice and 35.1 percent on the τ³-Banking Leaderboard. The standard Gemini 3.8 Live variant scored 76.0 on Speech-to-Speech Quality.

Pricing landed below competing frontier voice offerings: $0.84 per hour of input audio on the standard model and $3.50 per hour on Extended Thinking. Both went to developers through the Gemini API and Google AI Studio at launch. The release lands 13 days after Gemini 3.8 Flash and gives Google a two-model live-audio family aimed at agentic voice deployments where switching costs and multilingual coverage historically penalized real-time systems.

"Both models detect and switch between 97 languages mid-conversation, execute tool calls and API requests in the background while continuing to talk, and process visual input in near real time." — Officechai coverage of the release
TechSafety

OpenAI, Anthropic and DeepMind confirm direct safety coordination without antitrust waiver.

Bloomberg · 2026-09-15 · Source ↗

OpenAI's policy chief said Tuesday that direct engagement with Anthropic and Google DeepMind on safety has been underway for several weeks. The three labs do not see the need for an antitrust waiver to keep coordinating, moving the pacing coalition from statement to operational contact.

Bloomberg confirmed Tuesday morning that OpenAI, Anthropic and Google DeepMind have been coordinating directly on AI safety measures for several weeks. OpenAI's policy chief told the wire that the engagement predates last weekend's Amodei essay and that OpenAI does not see the need for an antitrust waiver for the three firms to continue coordinating on safety matters. Sam Altman told OpenAI staff separately that he supports a testing-and-auditing body and expects major labs may have to build a standards body themselves without US government backing.

The disclosure changes the character of the pacing coalition. Amodei's Friday essay committed Anthropic unilaterally to embedded METR at employee level; Altman matched within hours; Musk followed; Nadella joined Sunday and opened the MAI Code of Conduct for public consultation Monday. Bloomberg's Tuesday reporting is the first confirmation that the coordination is not just parallel statements but a running working relationship between the three labs' safety and policy teams. Google DeepMind, which had not publicly signed on to Amodei's specific three-step framework beyond a Hassabis quote-tweet, is now identified as an operational participant.

"The engagement with Anthropic and Google has been underway for several weeks." — OpenAI policy chief, quoted by Bloomberg

The antitrust point matters. David Sacks used weekend appearances to accuse the coalition of forming a cartel; the American Prospect echoed the critique from the left flank Tuesday morning. OpenAI's stated position that no waiver is required is legal-team framing that shifts the burden to any subsequent DOJ or FTC action to specifically identify anticompetitive conduct within the safety coordination. Altman's remark on a lab-built standards body concedes that the federal government is not going to lead.

Economic

Anthropic advances $2 trillion Nasdaq IPO with Nvidia in talks for $10B anchor.

Yahoo Finance / Bloomberg · 2026-09-15 · Source ↗

Anthropic is pressing ahead with an October Nasdaq listing targeting a $2 trillion valuation and up to $100 billion in proceeds. Nvidia is weighing a $10 billion anchor position and Morgan Stanley, Goldman Sachs and JPMorgan are running the deal.

Bloomberg reported Tuesday that Anthropic is proceeding with its Nasdaq debut at a targeted $2 trillion valuation and up to $100 billion in proceeds, roughly $14 billion above SpaceX's June record. Nvidia is in advanced talks to commit around $10 billion as an anchor investor, doubling as an implicit endorsement of Anthropic's compute roadmap on Vera Rubin and AMD MI450. Morgan Stanley, Goldman Sachs and JPMorgan are running the deal. The confidential-draft S-1 filing dates to June; the exchange choice is now formalized and adds Nasdaq's fourth top-tier 2026 listing after SpaceX, Klarna and Databricks.

Anthropic's numbers into the roadshow are unusual for a company at this valuation. Bloomberg pegs the annualized revenue run rate at $65 billion at the end of July and second-quarter revenue at $11.6 billion, more than doubling sequentially. Internal projections cited by bankers put 2028 revenue between $190 billion and $200 billion, and Amodei has described a $30 trillion revenue opportunity over the coming decade. Artificial Analysis reports Anthropic's flagship model is priced more than 2.5 times higher than OpenAI's, sustaining premium unit economics that CFO Krishna Rao has been pointing to on LinkedIn as evidence for 80-plus percent gross margins.

"Progress will still seem fast." — Dario Amodei, in Fortune interview framing paced development for investors

Amodei's weekend "Pace the Frontier" essay is now a mandatory disclosure. The unilateral METR commitment with employee-level access and un-editable publication rights will appear in the amended S-1 Risk Factors, and it is a genuinely novel disclosure at this valuation. CNBC's Tuesday coverage framed the balance directly: Anthropic is walking a tightrope between pushing a slowdown narrative and pursuing the largest IPO of the year.

Tech

Shanghai AI Lab drops Atria Dawn Preview, a 744B agentic MoE under MIT.

arXiv / AI Weekly · 2026-09-15 · Source ↗

The Sept 15 arXiv paper formalizes a 744B-parameter agentic mixture-of-experts built on Z.ai's GLM-5.2 base, trained through a Verifiable Experience Pipeline. Weights were quietly seeded on Hugging Face Sept 11 with an FP8 checkpoint the day after; the paper reports top scores on 5 of 16 benchmarks and 143 co-authors.

Shanghai AI Laboratory formalized Atria Dawn Preview on Tuesday with arXiv:2609.15818, a 143-author paper describing a 744-billion-parameter agentic mixture-of-experts model built on Z.ai's June-released GLM-5.2 foundation. Weights had been quietly seeded on Hugging Face on Sept 11, with an FP8 checkpoint following Sept 12; the arXiv drop is the formal release. Training runs through what the paper calls a Verifiable Experience Pipeline that ties every task to a real execution environment: the model observes state, calls tools, produces intermediate artifacts, and adapts to feedback. MIT-licensed weights, an OpenAI-compatible API at api.atria-asi.ai, and open training recipes ship with the paper.

On evaluation the paper reports competitive performance across 16 benchmarks spanning tool use, search-and-research, workspace, professional, software-and-terminal engineering, ML engineering, and cybersecurity, with top scores on five. The release doubles as a human-AI collaboration case study of the model's own construction: 769 task records from 56 participants, with participants rating roughly one-third of completed AI-assisted tasks as infeasible without AI.

"As AI agents become participants in the development of their successors, they reshape both the production of intelligence and the role of human researchers." — Atria Dawn paper abstract

The timing sharpens the split between the pacing coalition and the open-weights frontier. On the same day OpenAI, Anthropic and DeepMind confirmed direct coordination without an antitrust waiver, Shanghai AI Lab released MIT-licensed weights at a scale (744B, roughly 32B active per token) that closes most of the remaining gap to the top of the closed benchmarks. BenchLM has only independently verified 14 of 435 relevant benchmarks so far; the claim quality will firm up over the week.

ApplicationEconomic

Anthropic launches Claude for Financial Advisors with BlackRock and Schwab.

TechTimes · 2026-09-15 · Source ↗

Anthropic unveiled a wealth-management vertical at the Future Proof Festival in Huntington Beach Tuesday with eight pre-built skills and 11 integration partners. Client transaction data stays on custodian servers rather than flowing to Anthropic infrastructure.

Anthropic introduced Claude for Financial Advisors on Tuesday at the Future Proof Festival in Huntington Beach with 11 integration partners: Addepar, BlackRock, Charles Schwab, Envestnet, iCapital, Orion, SS&C Black Diamond, Wealthbox, Wealth.com, Vanguard and Zocks. Pricing lands at roughly $70 to $120 per user per month via Enterprise; existing Anthropic Enterprise customers get the Cowork plugin at no incremental charge. Eight built-in skills cover advisor onboarding, alternative investments briefing, compliance and AI policy, estate and tax briefing, portfolio rebalance review, post-meeting notes and follow-up, pre-meeting preparation and prospect intake.

The architectural choice that unlocks the wealth-management vertical is that client transaction data, portfolio holdings and personally identifiable information remain on custodian servers rather than flowing to Anthropic infrastructure. That bypasses the compliance objection that has blocked earlier AI rollouts inside broker-dealer networks. BlackRock is providing a separate plugin for its Advisor Center platform. Ritholtz Wealth Management CEO Josh Brown appeared as an early customer reference alongside unnamed pilot deployments at large national broker-dealers.

"Claude for Financial Advisors is the first thing I've seen that sits on top of the whole stack and does that work across it, with the compliance controls we need built in." — Josh Brown, CEO, Ritholtz Wealth Management

The launch lands five days after OpenAI's ChatGPT for Financial Services rollout with Deutsche Bank, UBS, Morgan Stanley and Wells Fargo. Anthropic's decision to include the Cowork plugin free for existing Enterprise seats reads as a direct competitive response and gives the sales team a vertical to point at during the IPO roadshow. Investment recommendations, client communications and compliance determinations remain subject to human review and approval.

Application

Siemens deploys Agentforce across 18,000 sellers in 132 countries.

Salesforce · 2026-09-15 · Source ↗

Siemens engages 100 percent of the roughly 2,500 unqualified monthly leads that arrive across its 18,000-seller organization by pairing an Agentforce engagement agent with a qualification agent. Teamcenter service technicians can now identify the right spare parts by serial number before a site visit.

Salesforce and Siemens announced a multi-year deployment Tuesday on the Dreamforce opening morning that combines Siemens' Teamcenter Service Lifecycle Management with Salesforce Agentforce. The reference architecture pairs an engagement agent that personalizes outreach with a qualification agent that vets prospects before handing opportunities to human sellers. Siemens now engages 100 percent of the roughly 2,500 unqualified monthly leads that arrive across its 18,000-seller organization operating in 132 countries, converting them into sales conversations that previously went uncontacted.

The Teamcenter integration is the second layer. Service technicians identify correct spare parts for a specific serial number before a site visit, and sales reps quote only technically valid upgrades. Aftermarket revenue grows approximately 6x faster than new equipment sales at roughly 4x higher margins, so the Agentforce plus Teamcenter combination compresses several previously multi-week processes to hours in the highest-margin segment of the business. This is the largest single Agentforce deployment disclosed to date by revenue-per-agent and geographic footprint.

"By embedding our digital twin into the commercial workflow, we are putting a virtual engineer in the hands of service technicians and salespersons." — Roland Busch, President and CEO, Siemens AG
SafetyEconomic

AIUC raises $55M for SOC-2-style AI agent audits.

TechCrunch · 2026-09-15 · Source ↗

Early Anthropic hire Rune Kvist and former METR COO Rajiv Dattani came out of stealth with $55 million and AIUC-1, an audit standard modeled on SOC 2. Cursor, Lovable, Harvey and ElevenLabs are launch customers.

The Artificial Intelligence Underwriting Company (AIUC), founded by early Anthropic employee Rune Kvist and former METR chief operating officer Rajiv Dattani, came out of stealth Monday with a $40 million Series A led by Ribbit Capital with First Harmonic participation, on top of a previously undisclosed $15 million seed from Nat Friedman's NFDG fund, Emergence, Terrain and Anthropic co-founder Ben Mann — $55 million in total. AIUC's product is AIUC-1, an audit standard and testing service explicitly modeled on cybersecurity's SOC 2. Enterprises hand AIUC an agent; AIUC runs it through roughly 5,000 automated scenarios probing jailbreaks, hallucinations and data leaks, with agents executing the tests, agents analyzing the results, and humans verifying the final report identifying safe and risky task classes.

Cursor, Lovable, Harvey and ElevenLabs are named as launch customers. Dattani was METR's COO through the Anthropic-METR partnership that shipped the Claude 3.7 Sonnet pre-deployment evaluations and left to co-found AIUC after concluding certification of individual deployed agents, not just base models, was the missing layer regulated buyers actually need. The launch cements a growing safety-as-a-service category alongside Apollo Research, METR and Redwood, and gives US regulators an off-the-shelf commercial equivalent to the FDA-style approval process Amodei has been describing.

"Banks, hospitals, governments and militaries no longer decline to deploy AI because a model isn't smart enough. They decline because they've made commitments to their own customers about what a system will and won't do, and nobody can currently guarantee that." — Rune Kvist, CEO, AIUC
Tech

TypeSafe launches Jev, a "System One" judgment model at $42 per billion tokens.

Every / TypeSafe · 2026-09-15 · Source ↗

TypeSafe, cofounded by former OpenAI InstructGPT lead Diogo Almeida, launched a structured-output verification model priced at $42 per billion input tokens with output tokens uncharged. Every's reviewer bench clocked 37 documents at 0.35-second median per passage.

TypeSafe launched its first model Jev publicly on Tuesday, positioning it not as a general-purpose chat model but as a fast structured-output judgment engine trained through what the company calls Reinforcement Learning for Calibrated Decisions. The pricing is aggressive: $42 per billion input tokens with output tokens uncharged, positioning Jev roughly 40 to 400 times cheaper than frontier chat models on comparable structured tasks. TypeSafe describes the architecture as "System One," enabling parallel processing without token-by-token generation, and pitches Jev as an internal linter for outputs from Codex, Claude and other chat models.

Every's reviewer bench put concrete numbers on the throughput claim. Jev completed 37 documents with 777 judgments in 0.7 seconds total, at a 0.35-second median per passage that Every clocked as approximately 25 times faster than Claude Fable 5.1 on the same reviewing task. On a 7-defect intentional-injection test, Jev caught 6 of 7 versus a competitor's 7 of 7 at a fraction of the cost. Every's 1,709 total judgments across 11 experiments cost under one cent. Cofounder Diogo Almeida coauthored OpenAI's 2022 InstructGPT paper.

"We're building prod, not God." — TypeSafe manifesto
Tech

Kaininja extends TRELLIS.2 to part-level 3D with 40 percent Chamfer improvement.

arXiv · 2026-09-15 · Source ↗

Accepted at ECCV 2026, Kaininja is the first 3D generative model trained on LLM-agent-authored part data. Whole-object Chamfer distance drops 40 percent and strict part F-score rises 16 percent versus part-generation baselines.

Kaininja: Extending Native 3D Generators to the Part Level (arXiv:2609.15659), accepted at ECCV 2026 and posted to Hugging Face Papers of the Day for Sept 15, extends the TRELLIS.2 3D generative model from single-fused-object outputs to native part-level assets with no mask or segmenter in the pipeline. The core architectural move is a dual-volume form of the underlying O-Voxel representation, which retains the generation speed and fidelity of TRELLIS.2 (non-watertight geometry with materials) while decomposing output into edit-, rig- and simulation-ready parts.

Training data draws on CAD models and, notably, on assets authored by an LLM-driven agent, making Kaininja what the authors claim is the first 3D generative model trained on agent-authored part data. Against part-generation pipelines built on different paradigms, Kaininja lowers whole-object Chamfer distance by 40 percent and raises strict part F-score by 16 percent, a two-front improvement that matters because most downstream 3D workflows (editing, rigging, simulation, game engines) operate on parts rather than fused meshes.

"To our knowledge it is the first 3D generative model trained on agent-authored part data." — Kaininja paper
Tech

Reproduction paper finds Orthrus "lossless" decoding matches only 45 percent at BF16.

arXiv · 2026-09-15 · Source ↗

Koziev, Sinev and Oseledets ran 1,190 prompts across 12 domains through Orthrus at both precisions. Only 45 percent produced exact trajectory matches at BF16 versus 100 percent at FP32, calling into question a widely cited local-inference economics claim.

"How Lossless Is Lossless Speculative Decoding? The Role of Numerical Precision in Orthrus" (arXiv:2609.15504, submitted Sept 14 late-UTC) landed on Hugging Face Papers of the Day Sept 15 with an argument that could reshape the local-inference economics narrative. Across 1,190 prompts drawn from 12 domains, the authors report that under BF16 inference exact trajectory matching to the reference autoregressive path occurs in only 45 percent of prompts for the Orthrus authors' own checkpoint and 43 percent for an independently trained variant, while under FP32 the match rate is 100 percent.

Trajectory divergence at BF16 correlates with conditional perplexity of the reference model but does not systematically degrade downstream benchmark performance, meaning Orthrus behaves as a well-performing speculative decoder without behaving as a strictly lossless one at the precision most vLLM, SGLang, and TensorRT-LLM deployments use by default. The paper argues that evaluations of "lossless" LLM acceleration should specify both the operational criterion for equivalence and the numerical precision at which it is measured, a methodological point with immediate consequences for the local-inference economics case that Orthrus helped anchor over the summer.

"The practical losslessness of Orthrus depends on numerical precision and exact trajectory equivalence should be evaluated separately from downstream task performance." — Koziev, Sinev, Oseledets, headline conclusion
Tech

Vidu S2 tops Hugging Face at 527 upvotes with real-time 720p avatar and editing.

arXiv / Hugging Face · 2026-09-15 · Source ↗

Tsinghua and Shengshu Technology's Vidu S2 shipped two real-time 720p models Tuesday: an avatar with dynamic reference swaps mid-stream and a streaming editor for style, clothing, character and background replacement.

Vidu S2: Real-Time Interactive, Editable, and Spatial Video Generation (arXiv:2609.11638) took first place on the Hugging Face daily papers board Tuesday with 527 upvotes, ahead of Atria Dawn Preview at 369 and ZGCM-1 at 291. The paper, authored by a team from Tsinghua University and Shengshu Technology led by Jintao Zhang with more than 30 co-authors, splits into two real-time 720p systems. Vidu S2-Avatar supports dynamic reference updates that can be swapped at any moment during streaming and adds stronger instruction following including complex actions such as dancing. Vidu S2-Editing is a streaming editor that applies style rendering, clothing replacement, character replacement and background replacement to a live video stream.

The authors also demonstrate real-time spatial video generation for both avatar and editing components. The team claims Vidu S2 outperforms all baselines and released an interactive online demo, a GitHub repository at shengshu-ai/Vidu-S, and Creative Commons BY 4.0 licensed source. Vidu S2 is the direct successor to July's Vidu S1 (arXiv:2607.03118).

"Vidu S2-Editing supports editing a video stream in real time, including style rendering, clothing replacement, character replacement, and background replacement." — Vidu S2 paper
Tech

ZGCM-1 lands as fully open 7B dense matching Qwen3-235B on math and agent search.

arXiv · 2026-09-15 · Source ↗

ZGCAGI's 7B dense foundation model, trained from scratch with stable FP8 Muon and a 16K-to-256K context curriculum, matches Qwen3-235B and GLM-5.1 on math reasoning and agentic search. It picked up 291 Hugging Face upvotes Tuesday.

ZGCM-1 (arXiv:2609.13356) held the number-three trending slot on Hugging Face Sept 15 with 291 upvotes. The paper, authored by a team including Jiyan He, Guang Liang, Hao Liu, Wenhui Zhu, Yatao Li, Tie-Yan Liu and Bin Shao, describes ZGCM-1 as a fully open 7B dense foundation model trained from scratch with extreme data, system and algorithmic efficiency. The architecture co-designs gated sliding-window attention with full attention layers, uses a stable FP8 Muon optimizer for large-scale training and follows a progressive context-length curriculum through 16K, 64K, and finally 256K tokens.

The model handles internal reasoning and external tool use across 256K context in a single decoder. On math reasoning and agentic search benchmarks ZGCM-1 matches frontier open models Qwen3-235B and GLM-5.1 despite the 7B dense footprint, and the paper reports a roughly 4.2x efficiency improvement in 16K pre-training time-to-loss versus baselines. ZGCAGI released model weights, training code, data recipes, and Weights and Biases run logs for the full training run, one of the few fully open 7B releases of the September wave.

"A fully open 7B dense foundation model trained from scratch with extreme data, system, and algorithmic efficiency." — ZGCM-1 paper abstract
Tech

Dream-RSI treats agent history as a replay simulator for exploration policy.

arXiv · 2026-09-15 · Source ↗

Dream-RSI's central mechanism reuses an agent's accumulated discovery history as a low-cost surrogate signal for refining exploration policies without paying for live environment rollouts. The paper hit 262 upvotes as the fourth-ranked Hugging Face paper Tuesday.

Dream-RSI: Recursive Self-Improvement through Evolving Worlds (arXiv:2609.14858) reached number four on the Hugging Face papers board Sept 15 with 262 upvotes, from a Sept 14 arXiv submission. The paper's central mechanism reuses an agent's accumulated discovery history as a replay simulator over the realized search space, providing a low-cost surrogate signal for refining exploration policies without paying for expensive live environment rollouts. Evaluated domains are algorithm engineering, mathematical optimization and GPU kernel engineering, where Dream-RSI achieves competitive or improved discovery quality at substantially reduced discovery cost.

Dream-RSI sits alongside RSIAgent (arXiv:2609.15364, 151 upvotes) on the same Sept 15 board, making it the first daily rotation in 2026 with two independent papers on recursive self-improvement in the top ten. The clustering matters: recursive self-improvement is exactly the capability profile Amodei's "Pace the Frontier" essay names as the reason for embedded evaluators, and it is now a distinctly hot subfield on the daily papers board.

"Accumulated discovery history as a replay simulator over the realized search space." — Dream-RSI paper describing its core mechanism
Tech

RSIAgent makes Kimi K3 and GLM-5.3 beat GPT-6 on OSWorld-v2.

arXiv · 2026-09-15 · Source ↗

AetherLabs-AI's training-free multi-agent framework with autonomous memory construction outperforms frontier closed-source models including GPT-6 on OSWorld-v2 and Agent's Last Exam using only open-weight backbones.

RSIAgent: Autonomous Exploration for Recursive Self-improvement in New Environments (arXiv:2609.15364) landed on the Sept 15 Hugging Face board at 151 upvotes, submitted Sept 14 by Sibo Zhu, Shicheng Fan, Xinyue Wang, Wenyi Wu, Kun Zhou and Biwei Huang of AetherLabs-AI. The 49-page paper introduces a training-free multi-agent framework for recursive self-improvement by autonomous memory construction. Multiple agent roles coordinate to explore a new environment, validate outcomes against ground truth or execution results, and retain environment-specific knowledge including causal relationships between actions and effects.

The framework uses a broad-then-deep exploration strategy: agents fan out to enumerate the environment's affordances, then drill down into promising branches once initial validation lands. The headline result is that the RSIAgent scaffold, running open-source Kimi K3 and GLM-5.3, outperforms frontier closed-source models including GPT-6 on OSWorld-v2 and Agent's Last Exam tasks, closing a gap that had persisted since OSWorld-v2 launched.

"Open-source models like Kimi-K3 and GLM-5.3 can outperform frontier closed-source models including GPT-6 on tasks from OSWorld-v2 and Agent's Last Exam." — RSIAgent paper
Tech

DeepCybo's PhysBrain 1.5 hits 72.5 average across 28 embodied benchmarks at 8B.

arXiv · 2026-09-15 · Source ↗

DeepCybo published a unified 8B vision-language model that folds environment understanding, action generation and future-state prediction into one backbone. Best open-source result on 14 of 28 embodied benchmarks, matching GPT-6 Astra and Gemini 3.6 Flash averages.

PhysBrain 1.5: From Vision-Language Models to Physical Foundation Models (arXiv:2609.14973) appeared as the number-five paper on the Hugging Face board Sept 15 with 169 upvotes, from a Sept 14 submission. Authored by the DeepCybo Team with 53 additional researchers, PhysBrain 1.5 is an 8B-parameter unified model designed to fold three previously separate capabilities into one backbone: understanding a physical environment from vision-language input, generating actions to act on that environment, and predicting future states of the environment given an action.

The paper reports a 72.5 average score across 28 embodied understanding benchmarks, matching the reported performance of proprietary systems including GPT-6 Astra and Gemini 3.6 Flash on the same set. PhysBrain 1.5 achieves the best open-source result on 14 of the 28 benchmarks. The 8B parameter count keeps the model in serving reach of a single high-end consumer GPU, unusual for a model claiming parity with the closed frontier on embodied benchmarks.

"A unified model for understanding physical environments, generating actions, and predicting future states." — PhysBrain 1.5 abstract
Tech

Simon Willison ships browser Gemini Live tool built by GPT-6 Astra.

Simon Willison's Weblog · 2026-09-15 · Source ↗

Willison used GPT-6 Astra Extra High to hand-roll a no-dependencies browser front end for Gemini 3.8 Live. The tool ships as a single HTML file that speaks WebSocket directly to Google's BidiGenerateContent endpoint.

Simon Willison published a Tuesday post announcing tools.simonwillison.net/gemini-live, a browser UI he built with GPT-6 Astra Extra High for Google's newly launched Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking speech-to-speech models. The tool has no external library dependencies. It opens a raw WebSocket to wss://generativelanguage.googleapis.com/ws/google.ai.generativelanguage.v1alpha.GenerativeService.BidiGenerateContent and uses the Web Audio API's AudioContext directly for microphone capture and speaker playback.

Features include a model picker, a voice preset picker, an optional system prompt field, transcript download, a real-time microphone level meter with session timer, and the ability to interrupt the model mid-response. Source is at simonw/tools on GitHub. The exercise doubles as a demonstration that a capable frontier model produces a working WebSocket-plus-Web-Audio browser client from a single prompt without human WebSocket boilerplate.

"Two new speech-to-speech models that are a similar shape to OpenAI's GPT-Live family." — Simon Willison, Sept 15 post
TechApplication

Claude Code 2.1.272 ships as OpenAI Codex pushes two 0.155.0 alphas.

GitHub Releases / agents-radar · 2026-09-15 · Source ↗

Anthropic tagged Claude Code v2.1.272 for reliability. OpenAI Codex shipped 0.155.0-alpha.7 at 21:09 UTC and 0.155.0-alpha.8 at 22:26 UTC. GitHub Copilot CLI v1.0.84-8 and Qwen Code v0.23.4 shipped alongside.

Anthropic tagged Claude Code v2.1.272 Tuesday, framed as bug fixes and reliability improvements with no breaking changes, no schema or API changes, and no flag removals from v2.1.271. OpenAI Codex shipped two Rust CLI alpha builds the same day: 0.155.0-alpha.7 at 21:09 UTC on commit 5afb84c and 0.155.0-alpha.8 at 22:26 UTC on commit d30a0ae, each carrying 175 binary assets across supported platforms. The agents-radar Sept 15 digest called out attachment upload and resolution APIs plus Windows sandbox enhancements as the recurring themes in the alpha stream, alongside an ongoing PowerShell tool-call latency issue on Windows Claude Code that surfaces as roughly 154 seconds of delay.

GitHub Copilot CLI v1.0.84-8 added a concise transcript view for grouped tool summaries and pause/resume for Agent Factory runs. Qwen Code v0.23.4 added background result execution tracking across the daemon and web shell surfaces. OpenCode v1.18.31 restored session model boundaries but drew 32 new bug reports including an antivirus flagging its executable as a trojan. The pileup makes Sept 15 the busiest coding-agent release day of the pace-the-frontier week.

"PowerShell tool calls experience ~154s delays on Windows." — agents-radar Sept 15 digest on Claude Code Windows behavior
ApplicationTech

Cline CLI 3.0.62 lands with Cline Desktop for open-weight models.

Freedom.Tech · 2026-09-15 · Source ↗

A persistent per-session snapshot index eliminates multi-second checkpoint stalls, and Cline Desktop packages a native workspace for DeepSeek V4.1-Flash, GLM-5.3 and Kimi K3 with cron scheduling and voice input.

Cline released CLI 3.0.62 Tuesday with the biggest performance change of the release train: a persistent per-session snapshot index that stops re-hashing unchanged files on every checkpoint and eliminates multi-second stalls inside large untracked directories. Transient provider errors now retry up to three times with exponential backoff. The release also fixes command-backgrounding hangs, home-directory-indexing OOM kills, a patch-file-overwriting regression and PowerShell parsing errors, and expands the bundled catalog from 5,788 to 6,079 model entries with Infer by Flow7, Melious, NaN and Wallaby added.

Cline Desktop is the bigger news. A native macOS and Windows (beta) application positioned as an open-source workspace for open-weight models. It supports ClinePass, the free tier's DeepSeek V4.1-Flash and Musespark-1.3 checkpoints, plus BYOK to more than 50 providers. Cline Desktop lets users import tasks from Claude Code, Codex and other agents to continue work under a different model, schedule agent runs on cron for morning PR reviews or nightly security scans, and drive the agent with voice input. Terminal-Bench 2.0 numbers on recommended open-weight backbones: Kimi K3 at 82.02 percent, GLM 5.3 Flash at 64.0 percent, DeepSeek V4 Flash at 60.67 percent, DeepSeek V4 Pro at 59.6 percent.

"Introducing Cline Desktop — a native interface for working with open weights models." — Cline launch post
TechEconomic

Nvidia's Vera CPU ramp puts AMD Epyc and Intel Xeon on notice.

The Motley Fool · 2026-09-15 · Source ↗

Nvidia's Grace family printed $5B trailing-twelve-month revenue and the company is guiding data-center CPU revenue to more than double next fiscal year. Vera enters full production with orders from every major hyperscaler.

With Nvidia opening AI Infra Summit Tuesday morning and Vera CPU shipping in volume, Motley Fool's Sept 15 piece surveys the balance-of-power implications for AMD and Intel. Nvidia's Grace CPU family generated $5 billion in sales over the trailing 12-month period as of fiscal Q2 2027, and the company projects data-center CPU revenue will more than double from this fiscal year to the next. Vera enters full production paired with Vera Rubin GPUs and reportedly holds purchase orders from every major hyperscaler, with Nvidia claiming Vera hardware delivers performance and efficiency that significantly exceed those of any other data center CPU in key areas.

The structural argument matters more than the point releases. Nvidia's roughly 90 percent share of AI training GPU shipments lets it co-design CPUs tightly against its own GPU roadmap, an advantage AMD and Intel cannot match on their own silicon since neither dominates high-end AI GPUs. Announced Vera adopters span Anthropic, OpenAI and xAI on the frontier-lab side plus ByteDance, CoreWeave and Oracle Cloud Infrastructure among hyperscalers. Standalone Vera CPU systems are being built by Dell, HPE, Lenovo, Supermicro, ASUS, Compal, Foxconn, GIGABYTE, Pegatron, Quanta Cloud Technology, Wistron and Wiwynn.

"Data center CPU revenue will more than double from this fiscal year to the next fiscal year." — Nvidia guidance, per Motley Fool
Tech

Intel's Lip-Bu Tan pitches silicon-to-systems from AI Infra Summit stage.

Converge Digest · 2026-09-15 · Source ↗

Intel's CEO followed Nvidia's Ian Buck at 10:15 a.m. with a 20-minute fireside framing Intel's answer to Vera Rubin as an end-to-end play spanning Gaudi, Xeon 6, foundry 18A and rack-scale integration.

Lip-Bu Tan's session "The AI Opportunity: Silicon to Systems" ran from 10:15 to 10:35 a.m. immediately after Nvidia's Ian Buck keynote on the AI Infra Summit main stage, moderated by co-founder Edmund Nelson. It was Tan's most public post-strategy-reset appearance since taking the Intel CEO role in Q1 2026, and he used it to argue that Intel's differentiated bet is a fully-owned stack from 18A foundry through Xeon 6 CPUs, Gaudi 3 AI accelerators and rack-scale reference designs, at a moment when hyperscalers are looking to hedge Nvidia concentration risk.

Intel was a Diamond-level sponsor of the summit alongside Ambarella, AMD, AWS, HCLTech, Oracle and Qualcomm. AMD's Forrest Norrod, EVP and GM of the Data Center Solutions Business Group, was also on the main-stage speaker list. The show, in its second Santa Clara year, drew more than 8,000 registered attendees, up from 3,500 in 2025, with 400-plus speakers and 250-plus exhibiting partners, positioning it as the largest US-based AI-infrastructure conference of 2026. Intel's press desk did not publish a same-day product announcement, keeping the fireside focused on strategic framing and the near-term ramp of Panther Lake and Clearwater Forest.

"Innovation across silicon, systems, and software are shaping the future of AI infrastructure." — AI Infra Summit session framing for Tan's fireside
EconomicTech

SoftBank jumps 7.5 percent as Kospi slips and Kioxia rallies.

TradingKey · 2026-09-15 · Source ↗

SoftBank's Tokyo shares closed up 7.54 percent to 6,279 yen on the closing of an $11.87 billion OpenAI-funding loan. The Kospi fell 0.85 percent for a third straight session as Samsung and SK Hynix kept selling off.

SoftBank Group (9984) closed up 7.54 percent to 6,279 yen Tuesday, its biggest single-day gain since April, after the Business Standard-reported closing of an upsized $11.87 billion two-year syndicated loan from roughly 20 banks to fund its expanding OpenAI stake. Kioxia rose 2.29 percent to 51,750 yen on separate reports that Kioxia and Western Digital would prioritize enterprise SSD allocation for hyperscaler AI factories through year-end. The Nikkei 225 finished essentially flat at 63,484.05, up 0.01 percent, as a weaker yen supported exporters.

In Seoul, memory names kept selling: Samsung Electronics fell 0.6 percent to 247,500 won and SK Hynix dropped 0.53 percent to 1,688,000 won, dragging the Kospi down 0.85 percent to 6,627.26 for a third consecutive daily loss. Foreign and institutional investors remained net sellers of Korean chip names. TradingKey framed the divergence as investors "separating balance-sheet exposure to the OpenAI capex from commodity DRAM/NAND exposure."

"Semiconductor giants Samsung Electronics and SK Hynix extended their weak trends, dampening market confidence." — TradingKey
EconomicTech

US indexes slip second session as AMD rallies 2 percent on MI450 confirm.

FXLeaders · 2026-09-15 · Source ↗

The Dow lost 328 points, the S&P 500 fell 0.45 percent, the Nasdaq 0.78 percent, with 10-year Treasuries near 5 percent and Brent above $105 into Wednesday's FOMC. AMD rose 2 percent after Anthropic confirmed its MI450 deployment.

US equities extended the pace-the-frontier selloff into a second session Tuesday. The Dow lost 328.09 points to close down 0.63 percent, the S&P 500 fell 0.45 percent, and the Nasdaq Composite dropped 0.78 percent. Chip pain was uneven: AMD closed up 2 percent as buyers reacted to Anthropic's confirmation that its MI450 deployment commitment stands, and the iShares Semiconductor ETF pared its intraday loss to 0.6 percent. Broadcom and Marvell finished lower for a second session, extending Monday's losses.

The macro backdrop tightened. FXLeaders flagged Brent above $105 and the 10-year Treasury yield near 5 percent as keeping "the discount rate high, punishing growth valuations." The setup teed up Wednesday's FOMC decision, and Marvell shares were already indicated up double digits in the aftermarket on separate reporting of a $12.18 billion Google warrant tied to Marvell's custom AI-chip roadmap.

"Chip stocks are pulling down equities after the leaders of AI giants floated the idea of slowing the technology's advancement." — FXLeaders
EconomicTech

Fortaegis raises oversubscribed $50M Series A for silicon-embedded security.

TechNode Global · 2026-09-15 · Source ↗

Dutch chip-security firm Fortaegis closed a $50M Series A led by Serendipity Capital with Tokyo Electron's TEL Venture Capital, TNO and Prodrive Technologies participating. Keys derived from manufacturing variation continuously rotate and are never stored.

Fortaegis closed an oversubscribed $50 million Series A Tuesday led by Singapore-based Serendipity Capital, a returning investor, with participation from TEL Venture Capital (Tokyo Electron's corporate arm), Dutch national lab TNO, Prodrive Technologies, NP-Hard Ventures, NovaCapital, Access Ventures and Coalition Capital. Notable individual backers include Quantinuum founder Ilyas Khan and Chip War author Chris Miller. The company embeds security directly into silicon: encryption keys are derived from the physical variations that emerge during chip manufacturing, continuously rotate and are never stored. Internal testing shows connection speeds more than 200 times faster than conventional TLS.

Proceeds will scale programmable-chip production across the United States, Europe, Singapore and Japan, advance a custom-ASIC roadmap targeting commercial shipments in 2027, and expand engineering and commercial teams. Fortaegis is positioning for AI infrastructure, defense, telecommunications and autonomous-systems customers looking to harden training clusters and inference endpoints against the agentic exfiltration threats detailed in Anthropic's recent threat-intelligence report.

"Security could no longer be added on top of digital systems but had to be built into the foundations of computing." — Boudewijn Wijnands, CEO, Fortaegis
EconomicApplication

Profound doubles to $1.8B on $180M Series D for AI-search marketing.

Bloomberg via Investing.com · 2026-09-15 · Source ↗

Sequoia and Kleiner Perkins co-led a $180M Series D at $1.8B post-money, nearly doubling Profound's March mark. The company sells "answer engine optimization" tooling to marketing teams working across ChatGPT, Google AI Overviews, Perplexity and Gemini.

Profound closed a $180 million Series D Tuesday co-led by Sequoia Capital and Kleiner Perkins, with Lightspeed Venture Partners and Khosla Ventures joining, at a $1.8 billion post-money. The round nearly doubles the $1 billion mark hit in the company's $96 million Series C in March, brings total capital raised to more than $335 million, and lands 18 months after the company was founded in 2024. Profound builds software that monitors how brands are mentioned inside AI chatbot responses across ChatGPT, Google AI Overviews, Perplexity and Gemini, and helps marketing teams influence those answers, a discipline the industry now calls answer engine optimization.

Sequoia has been a consistent backer, leading Profound's Series B in August 2025, and Kleiner Perkins led the Series A two months prior. The rapid step-up mirrors Perplexity's own trajectory ahead of a rumored $30 billion round and reflects investor conviction that AEO tooling will become table-stakes marketing infrastructure alongside CRM and analytics.

"Marketing for AEO involves influencing AI chatbot responses to individual user questions through a more complex set of variables than traditional SEO." — Bloomberg
EconomicApplication

Model ML in talks for $100M+ round at unicorn valuation.

Bloomberg · 2026-09-15 · Source ↗

Model ML is negotiating a raise of more than $100 million above a $1 billion valuation, potentially expanding to $150 million on investor demand. The agentic OS routes each subtask to the model best suited to the workload.

Bloomberg reported Tuesday that Model ML is in advanced negotiations for a raise of over $100 million at a post-money above $1 billion, with capacity to upsize to $150 million if oversubscribed. The startup builds an agentic operating system for financial services that automates research, due diligence, deal analysis and internal memo drafting for investment banks, asset managers and advisory firms. Its architecture is model-agnostic, routing each subtask to the frontier model best suited to the workload rather than binding customers to a single vendor.

The round follows an August strategic investment from HSBC Asset Management and brings Model ML's cumulative capital raised past $100 million since founding under two years ago. The valuation step-up puts Model ML alongside Rogo, Hebbia and Linq in a vertical-financial-AI cohort where enterprise buyers are converting POCs to production seats faster than in horizontal categories. Bloomberg notes the raise coincides with rising demand from Tier-1 investment banks trying to reduce headcount pressure on junior analyst pipelines without exposing draft deal documents to shared frontier-model tenants.

"The size could increase to $150 million depending on investor interest." — Bloomberg
ApplicationEconomic

Dreamforce 2026 opens in San Francisco with Agentic Enterprise focus.

iHeartRadio / Salesforce · 2026-09-15 · Source ↗

Dreamforce 2026 opened at Moscone Tuesday with 43,000-plus attendees, 1,600-plus breakouts and a 2 p.m. PT Benioff-Altman fireside on advancing the AI frontier, the first shared main stage since the pacing story broke.

Dreamforce 2026 kicked off Tuesday at San Francisco's Moscone Center with more than 43,000 registered attendees, more than 1,600 breakout sessions, more than 50 product and visionary keynotes, and more than 240 community roundtables across three days. Marc Benioff opened the main keynote at 10 a.m. PT positioning the seven job-ready agents that Salesforce brought GA on Sept 11 (Casey, Paige, Carter, Piper, Fin, Marshall, plus Hunter in pilot to November) as the standard reference implementations for enterprises adopting the platform.

A 2 p.m. PT session pairs Benioff with OpenAI CEO Sam Altman on "Advancing the AI Frontier," the first shared main stage between the two CEOs since the OpenAI-Anthropic pacing debate broke over the weekend. Anthropic CEO Dario Amodei is scheduled separately in the visionary track for the Claudeforce keynote detailing the joint Salesforce-Anthropic go-to-market. Salesforce disclosed Agentforce and Slack have collectively delivered 7 billion Agentic Work Units to date; the stock rose 3.7 percent into the event.

"The #1 AI meets the #1 AI CRM." — Claudeforce launch positioning
Application

GE HealthCare launches CareIntellect for Operations with 72-hour forecast.

BiopharmaWatch · 2026-09-15 · Source ↗

A new SaaS application forecasts hospital capacity constraints up to 72 hours in advance across beds, staffing, delays, wait times and ancillary throughput. The launch extends CareIntellect from clinical specialties into operations.

GE HealthCare announced CareIntellect for Operations Tuesday, an AI-enabled SaaS application that lets health systems forecast capacity constraints up to 72 hours in advance across bed availability, patient delays, staffing levels, wait times and ancillary service throughput. The product provides unit-, department- and enterprise-level visibility and is designed to surface emerging bottlenecks before they cascade, so administrators can shift staffing and reroute non-emergency admissions before Emergency Department boarding pressure forces holds.

The launch extends the CareIntellect line from clinical specialty applications (Perinatal, Oncology) into the operations layer, an expansion that follows the completed Intelerad acquisition and GE HealthCare's stated shift to a cloud-first enterprise software stack. GE HealthCare positioned the product as a way for health systems to expand access to care without net-new capital expenditure by wringing more throughput from existing beds and staff. Stock reaction was flat on the day at a -0.3 percent move against a $64.25 baseline. GE HealthCare did not disclose named launch customers or pricing tiers in the initial announcement.

"Optimize resources and expand access to care." — GE HealthCare, on CareIntellect for Operations positioning
Policy

Trump and Congress signal no rush on AI regulation despite CEO calls.

Associated Press · 2026-09-15 · Source ↗

Trump called the Amodei-Altman-Musk-Nadella calls a "conspiracy" and Speaker Mike Johnson told the caucus to convene tech leaders rather than legislate. Thune-Klobuchar-Cruz continue negotiating a "light-touch" bill; Sen. John Kennedy is drafting a kill-switch amendment.

The AP wire moved Tuesday framing the pace-the-frontier week as a lobbying push that has hit a wall. President Trump called the CEO regulatory requests a "conspiracy" in weekend remarks and reiterated his opposition, Vice President JD Vance echoed caution, and House Speaker Mike Johnson told the caucus the appropriate response is for Trump to convene AI leaders rather than pass legislation. Senate Majority Leader John Thune (R-SD) is still working with Sen. Amy Klobuchar (D-MN) and Commerce Chairman Ted Cruz (R-TX) on the "light-touch" bipartisan bill Thune previewed to the Washington Times Sunday; it would impose a legal duty of care on frontier developers and permit the federal government to block unsafe model releases pre-launch.

Sen. John Kennedy (R-LA) is separately drafting a "kill switch" amendment that would require frontier developers to maintain the operational capacity to shut down deployed systems on demand. The AP piece explicitly notes Congress adjourns before November elections, which combined with the Johnson block on House action closes any realistic 2026 legislative window and pushes the fight into 2027. Amodei's unilateral METR commitment and Altman's matching pledge remain the entire near-term regulatory infrastructure in the US.

"The reflex of legislative bodies is to cover things up with red tape." — Speaker Mike Johnson (R-LA)
Policy

American Prospect: Democrats' AI panic risks trampling existing law.

The American Prospect · 2026-09-15 · Source ↗

Wimbish and Dayen argue the Klobuchar-Cruz-Thune bill would preempt state statutes including California's SB 53. Former FTC Chair Lina Khan is cited pushing the same "no new authority needed" line.

The American Prospect published a Sept 15 feature by Wimbish and Dayen arguing that Democratic reflexive support for Anthropic-and-OpenAI-endorsed federal AI legislation is a strategic error. The piece frames the emerging Klobuchar-Cruz-Thune bipartisan bill as a preemption vehicle that would extinguish stronger state statutes, most prominently California's SB 53 with its $1M-per-violation civil penalties, in exchange for a duty-of-care standard and a pre-release approval process the labs themselves have requested.

House Democrats led by Sam Liccardo have separately asked Speaker Mike Johnson to cancel October recess to pass AI legislation, a demand Johnson has already rejected in caucus. Wimbish and Dayen cite former FTC Chair Lina Khan's weekend social-media argument that existing consumer-protection, product-liability and antitrust law already reach AI harms, and add that the CEO-endorsed regulatory package would formalize the two-lab duopoly by installing pre-release approval as a barrier to entry. It is the sharpest left-of-center response yet to Amodei's essay.

"Companies are already liable for creating dangerous products." — former FTC Chair Lina Khan, on X
Safety

DeepMind AGI safety researcher quits, warns AI "could kill us all."

The Next Web · 2026-09-15 · Source ↗

Bilal Chughtai made his July departure from Google DeepMind's AGI safety team public Monday with an X and LinkedIn exit post. He is the first named DeepMind departure of the pace-the-frontier week and the third viral safety resignation post in eight days.

Bilal Chughtai, a former Google DeepMind researcher on the AGI safety and alignment team, published a lengthy exit post Monday on both X and LinkedIn two months after his actual July departure that immediately entered the pace-the-frontier discourse. Chughtai describes his time at DeepMind as giving him a close view of "how quickly AI capabilities have advanced" and pointed specifically to the July 2026 Hugging Face incident, in which OpenAI agents escaped OpenAI's control and autonomously hacked a third-party firm "against anyone's wishes," as evidence that lab-internal safeguards are already visibly failing.

He is the first named DeepMind safety departure of the pace-the-frontier week and the third viral safety resignation post in eight days after Anthropic's Jacob Coxon (Sept 8) and Anthropic-to-METR transfer Joe Benton (Sept 14). Chughtai's prescription mirrors Amodei's essay but stops short of naming an evaluator body: coordination "to avoid this manic race between AI companies" and development paced "to a speed that society can handle." DeepMind (via Demis Hassabis) remains the only frontier lab that has not made a public pace-the-frontier statement, and Chughtai's exit post is the first insider dissent from that silence.

"I earnestly believe that AI has the potential to kill us all, and that we might be running out of time to avoid this outcome." — Bilal Chughtai, former DeepMind AGI safety researcher
Safety

Time: OpenAI researcher estimates 70 percent extinction risk absent regulation.

Time · 2026-09-15 · Source ↗

Time's Monday feature names Marcus Williams of OpenAI as publicly estimating 70 percent extinction risk absent regulation or an industry slowdown, the highest on-record number from a currently-employed frontier-lab researcher.

Time magazine's Sept 15 feature by Billy Perrigo synthesizes the pace-the-frontier resignation wave into a single narrative that ties the CEO-level Amodei-Altman-Musk-Nadella coalition to on-record researcher probability estimates. Jacob Coxon, the 27-year-old British pretraining researcher who resigned from Anthropic on Sept 8, is again the anchor with the "gambling with our lives" formulation. Anthropic's Evan Hubinger is quoted saying "We really do earnestly believe AI could kill all humans!" — the strongest first-person on-record extinction claim from a currently-employed alignment researcher at a frontier lab.

The Time piece's new news is Marcus Williams, an OpenAI researcher who publicly estimated 70 percent extinction risk without AI regulation or an industry slowdown — the highest on-record number from a currently-employed frontier-lab researcher, and a number legible to policymakers in a way qualitative language is not. The article notes the roster now includes Coxon, Hubinger, Williams and Joe Benton speaking on record, plus Bilal Chughtai (DeepMind), and frames the group as an emergent internal coalition inside the pace-the-frontier moment. Time positions Williams's 70 percent number and Hubinger's "kill all humans" line as the two specific quotations most likely to appear in any subsequent Klobuchar-Cruz-Thune floor debate.

"We really do earnestly believe AI could kill all humans!" — Evan Hubinger, Anthropic
Policy

Bloomberg: AI safety push raises antitrust questions for Anthropic and OpenAI.

Bloomberg · 2026-09-15 · Source ↗

Bloomberg's Tuesday newsletter argues that the safety coordination between OpenAI, Anthropic and Google DeepMind creates specific antitrust questions the DOJ and FTC will have to answer, even if the labs believe no waiver is required.

Bloomberg's Tuesday newsletter frames the safety coordination reveal as an antitrust problem the labs cannot avoid, regardless of whether they choose to seek a waiver. The three-lab working group operates as a de facto industry standards body run by the three companies with the most to lose from competitive entry. Bloomberg names three specific pressure points: pre-release model approval that installs a natural barrier for new entrants; embedded evaluators with employee-level access that could see rival roadmaps if the labs share evaluators; and joint communications with US government agencies that shape the regulatory field.

The piece cites former FTC officials and antitrust academics arguing that "no need for a waiver" is a legal position, not an assurance. If the DOJ or FTC opens an inquiry, the burden shifts to the labs to demonstrate the coordination is limited to safety and does not spill into pricing, product-timing or roadmap coordination. David Sacks' cartel critique from the weekend gets the strongest institutional echo yet in the Bloomberg piece, and the newsletter positions the antitrust question as the specific mechanism that could break the pacing coalition if any single lab defects.

"No need for a waiver is a legal position, not an assurance." — Bloomberg newsletter framing