Daily2026-09-06
PolicySafety

Anthropic pays $1.5 billion to authors — largest copyright recovery in US history

Anthropic filed a proposed class-action settlement Friday September 5 in the Northern District of California in Bartz v. Anthropic, agreeing to pay at least $1.5 billion — roughly $3,000 per book across an estimated 500,000 works — for downloading pirated titles from LibGen and PiLiMi (Pirate Library Mirror) to train Claude. Judge William Alsup set preliminary approval for September 8, and Anthropic will destroy the pirated datasets across four installment payments through September 2027. The payout functions as a per-book price floor for every remaining AI training-data lawsuit and lands hours before Grady Hendrix and Jennifer Roberson filed the 49th US AI copyright case against Apple over Books3 training of OpenELM, represented by the same Susman Godfrey firm.

Policy · Safety · Bloomberg / NPR / Authors Guild · 2026-09-05 · Source ↗

Anthropic on Friday September 5 filed a proposed class-action settlement in the Northern District of California in Bartz v. Anthropic committing to pay at least $1.5 billion to resolve author claims that it downloaded pirated books from Library Genesis and Pirate Library Mirror to train Claude. The payment works out to roughly $3,000 per book across an estimated 500,000 titles, and it is the largest publicly reported copyright recovery in United States history. Class representatives are thriller novelist Andrea Bartz and nonfiction writers Charles Graeber and Kirk Wallace Johnson. Judge William Alsup set the preliminary-approval hearing for September 8 in the same courthouse where he had ruled in June that training Claude on lawfully purchased books can qualify as fair use but that sourcing the training corpus from shadow libraries cannot.

The settlement is structured as four installments over two years: October 2, 2025; April 30, 2026; September 25, 2026; and September 25, 2027. Anthropic will also destroy the pirated datasets. The payout is Anthropic's alone; the fair-use pathway that Judge Alsup blessed in June continues to protect Claude's training paradigm on legally acquired books going forward. The company's counsel described the payment as covering "legacy claims" that predate its shift to licensed and purchased data.

What the number does to the rest of the industry

The settlement lands as a per-book price floor for every remaining author copyright case pending against a frontier lab. OpenAI, Meta, Microsoft, Cohere, Perplexity, Nvidia, Databricks, Suno, Udio, Stability and, as of Friday afternoon, Apple are named defendants in some form of AI copyright complaint filed since the New York Times v. OpenAI action of December 2023. The Susman Godfrey firm that led the Anthropic case also filed the Apple complaint on behalf of horror novelist Grady Hendrix and fantasy author Jennifer Roberson in the same Northern District of California hours after the Anthropic term sheet reached the docket. Both cases target the Books3 pirated corpus and both seek statutory damages, disgorgement and destruction of the infringing models — the template Judge Alsup approved in the Anthropic settlement.

The Hendrix and Roberson complaint against Apple is the 49th US copyright lawsuit against an AI company and the 18th brought by book authors. Justin Nelson, the Susman Godfrey partner who led the Anthropic negotiation, told reporters after filing that "each of these cases raises the same fundamental question of who gets paid when a model learns from a book." The commercial read is that the class-action bar has now systematically mapped every major frontier lab against the same LibGen/PiLiMi source-of-truth and priced each infringing corpus at Alsup's $3,000-per-work benchmark.

"AI companies cannot simply steal authors' creative work to build their AI products and profit off of them." — Mary Rasenberger, Authors Guild CEO

What the number does to Anthropic

The $1.5 billion payout is roughly 3.5 percent of the $45 billion implied annual revenue run rate that Anthropic disclosed to Series F investors last week during its $183 billion valuation announcement. It is comfortably within the war chest but the timing matters: the settlement filed the same week Anthropic told investors it is on a fall IPO track sets a public expectation that similar payouts to other rights-holders will be part of the business through 2027. Anthropic paired the filing with a public commitment that its current training data pipeline sources from licensed publishers, purchased physical books scanned in-house, and synthetic data only.

"We remain committed to building safe AI systems that help people and organizations extend their capabilities, advance scientific discovery, and solve complex problems." — Aparna Sridhar, Anthropic Deputy General Counsel

The larger frame

Judge Alsup's June ruling had already split the frontier training-data question into two rulings the industry accepted: buy-and-scan is fair use; shadow-library scraping is not. Friday's settlement puts a dollar figure on the second half of that ruling. The larger implication for the AI industry is that pirated-corpus liability is now finite and priced. Every remaining defendant can now model exposure on its own case as "approximately $3,000 per book identified in a shadow library it can be shown to have pulled from," and the ambiguity that had made the class-action risk uncappable for eighteen months has resolved to a number.

Authors Guild CEO Mary Rasenberger, whose organization has led the plaintiff-coordinating effort across the 49 pending cases, called it "meaningful compensation" and framed it as the moment where AI companies "cannot simply steal creative work." Anthropic board chair Bret Taylor, who is separately being pressured by California Attorney General Rob Bonta and Delaware AG Kathleen Jennings on OpenAI child-safety issues the same day, did not comment on the settlement. The preliminary approval hearing on September 8 is expected to be routine.

"This landmark settlement will provide meaningful compensation for each class work and sets a precedent of AI companies paying for their use of pirated books." — Justin Nelson, Susman Godfrey partner and lead class counsel
Big Picture

The class-action bar priced pirated AI training data on Friday. Anthropic agreed to pay $1.5 billion at roughly $3,000 per book across 500,000 works, the largest US copyright recovery on record, and hours later the same Susman Godfrey firm filed the 49th AI copyright case against Apple over the Books3 corpus. Every remaining defendant now models exposure against Judge Alsup's number. The buy-and-scan pathway the court blessed in June remains the safer route forward.

Trump hosted 33 tech executives at the White House the night before as Mark Zuckerberg pledged at least $600 billion in US investment through 2028, Tim Cook committed $600 billion and Sundar Pichai committed $250 billion over two years. Elon Musk was pointedly excluded. Broadcom Q3 printed $15.95 billion of revenue with $5.2 billion of AI-semiconductor revenue up 63 percent, guided Q4 AI to $6.2 billion, and disclosed a fourth $10 billion custom-XPU customer analysts identified as OpenAI, sending AVGO up 11 percent and NVDA down 2.5 percent. US August payrolls added just 22,000 jobs versus 75,000 consensus with June revised to negative 13,000, taking unemployment to 4.3 percent as Salesforce's Marc Benioff confirmed cutting customer support headcount from 9,000 to 5,000 through Agentforce.

California AG Rob Bonta and Delaware AG Kathleen Jennings met OpenAI's legal team over the Adam Raine wrongful-death suit and the Old Greenwich murder-suicide, warning safeguards "did not work." OpenAI released "Why Language Models Hallucinate" arguing benchmarks reward confident guessing. Moonshot AI shipped Kimi K2-0905 at 1T parameters with 256K context under modified MIT license, Groq serving it at 445 tokens per second. Google released EmbeddingGemma at 308M parameters for on-device multilingual RAG. Anthropic Research announced Claude produced the first end-to-end Lean proof of Fermat's Last Theorem across 11 days of autonomous work on the Prove2Me platform, 13 million lines and 29,500 intermediate theorems. Fidji Simo launched the OpenAI Jobs Platform with a 10 million US certifications target and Walmart, BCG, John Deere and Accenture as partners. Baseten closed $150 million Series D at $2.15 billion. Tesla's Robotaxi iOS app crossed 2 million downloads on launch day. Samsung debuted Vision AI Companion at IFA Berlin and Neura Robotics put its 4NE1 humanoid on the floor. Pennsylvania Act 35 criminalizing forged AI likenesses took effect. WSJ exposed US chip diplomacy dangling Nvidia Blackwell access to Armenia to close the Azerbaijan peace deal. Warp launched Warp Code, Chris Lattner made the case for Mojo, and Simon Willison christened GPT-5 Thinking as "Research Goblin."

More today

24 stories · 2026-09-04 — 2026-09-05
EconomicPolicy

Trump hosts 33 tech CEOs at White House as pledges pass a trillion dollars

Deseret News / CBS News · 2026-09-04 · Source ↗

Mark Zuckerberg pledged at least $600 billion of US investment through 2028, Tim Cook committed $600 billion, Sundar Pichai committed $250 billion over two years, and Satya Nadella cited $75-80 billion of FY-annualized AI infrastructure at a State Dining Room dinner staged as "American AI Dominance." Elon Musk was pointedly excluded from the guest list; a Tesla representative attended in his place.

President Trump hosted 33 people at a White House dinner Thursday September 4 with a guest list that read as a roll-call of frontier AI: Mark Zuckerberg (Meta), Tim Cook (Apple), Sundar Pichai and Sergey Brin (Google/Alphabet), Satya Nadella (Microsoft), Sam Altman and Greg Brockman (OpenAI), Lisa Su (AMD), Safra Catz (Oracle) and Bill Gates. Elon Musk and Jensen Huang did not attend; a Tesla representative filled Musk's chair. The staged theme was "American AI Dominance," and Trump asked each executive around the table to state what their firm plans to spend in America.

The pledges landed near $1.5 trillion in aggregate. Zuckerberg publicly upgraded Meta's US investment commitment to "at least $600 billion" through 2028, up from the $300 billion he had said publicly before. Apple committed $600 billion, Google pledged $250 billion over the next two years, and Microsoft cited its $75-80 billion FY-annualized AI infrastructure spend. Sam Altman, who told the room the change from prior administrations was "very refreshing," used the dinner to announce the OpenAI Jobs Platform to the president directly. First Lady Melania Trump used the same evening to launch an AI education initiative aimed at K-12 classrooms.

"We're going to be investing something like at least $600 billion in the U.S. through 2028." — Mark Zuckerberg, CEO of Meta

Beyond the pledges, the optics did work. The dinner formally aligned the largest US frontier labs and hyperscalers with the White House against Chinese competition and it publicly punished Musk for his rift with the administration. The trillion-dollar-plus capex commitments through 2028 also compound the compute-buildout math that Broadcom and Anthropic disclosed the same week: hyperscalers are now publicly bound to a spending trajectory that dwarfs any prior industry cycle.

EconomicTech

Broadcom Q3 AI revenue up 63% to $5.2B, guides Q4 AI to $6.2B on fourth $10B custom-XPU customer

CNBC / Yahoo Finance / Broadcom IR · 2026-09-04 · Source ↗

Broadcom printed $15.95 billion of total revenue up 22 percent year over year with $5.2 billion of AI-semiconductor revenue up 63 percent, guided Q4 AI to $6.2 billion, and Hock Tan disclosed a fourth $10 billion custom-XPU commitment that analysts identified as OpenAI. AVGO closed up 11 percent Friday adding roughly $150 billion of market cap, while NVDA dropped 2.5 percent on the read that hyperscaler diversification into custom silicon is now a durable trend.

Broadcom reported fiscal Q3 2025 after the bell Thursday September 4, printing $15.95 billion of total revenue up 22 percent year over year and $5.2 billion of AI-semiconductor revenue up 63 percent — an eleventh consecutive quarter of AI-segment growth and the fifth in a row where AI grew faster than 60 percent. Semiconductor Solutions revenue landed at $9.2 billion up 57 percent with AI accelerators (Google TPU, Meta MTIA, ByteDance) and networking silicon (Tomahawk, Jericho) as the two biggest drivers. Infrastructure Software, which houses VMware, was $6.8 billion up 43 percent.

Hock Tan told the call that a fourth $10 billion custom-XPU customer had placed production orders that will hit Broadcom's revenue in fiscal 2026, and Q4 AI revenue would land at $6.2 billion — above the $5.8 billion Street consensus. Ben Reitzes at Melius Research, Stacy Rasgon at Bernstein and Harlan Sur at JPMorgan all identified the new customer as OpenAI on Friday morning notes, citing OpenAI's public shift toward custom silicon for inference workloads.

"We believe an incremental $10bn order for AI Racks over 3 quarters coincides with the timing of a new prospective XPU partner (Open AI)." — Ben Reitzes, Melius Research analyst

The market reaction on Friday split the two names. AVGO climbed roughly 11 percent, adding about $150 billion of market cap in a single session and vaulting Broadcom into the top five most valuable US companies. NVDA closed down 2.5 percent as investors weighed whether OpenAI's diversification erodes the Blackwell dominance narrative. Rasgon called Broadcom "the leading alternative" to Nvidia in accelerator silicon, taking the position previously held by Marvell. Total AI-semi revenue is now roughly a third of Broadcom's consolidated revenue.

EconomicTrend

US August payrolls add 22K, unemployment climbs to 4.3% — weakest AI-cycle print

CNBC / Bureau of Labor Statistics · 2026-09-05 · Source ↗

Payrolls grew just 22,000 versus the 75,000 consensus, June revised to negative 13,000 (first negative month since December 2020), and the U-3 unemployment rate ticked to 4.3 percent — its highest since October 2021. The soft print landed hours after Marc Benioff confirmed cutting Salesforce customer-support headcount from 9,000 to 5,000 through Agentforce automation.

The Bureau of Labor Statistics released the August employment situation Friday morning September 5. Nonfarm payrolls grew 22,000 against a 75,000 consensus, the June print was revised to a loss of 13,000 (the first negative month since December 2020), and the U-3 unemployment rate ticked up to 4.3 percent, its highest since October 2021. Broader U-6 unemployment climbed 0.2 points to 8.1 percent. Health care added 31,000 jobs and social assistance added 16,000; wholesale trade and manufacturing each shed 12,000. Total year-to-date monthly job growth is now running at 61,000 versus the pre-pandemic eighty-year average of roughly 150,000.

The soft print landed hours after Salesforce CEO Marc Benioff publicly confirmed on The Logan Bartlett Show that customer-support headcount had dropped from 9,000 to 5,000 through Agentforce automation. Palantir, Amazon and Meta all telegraphed further AI-driven workforce restructuring during the week. Economists at Goldman Sachs, Morgan Stanley and Bank of America moved to a 50-basis-point cut at the September 17 FOMC as base case within minutes of the release; 2-year Treasury yields dropped 12 basis points on the print.

"I've reduced it from 9,000 heads to about 5,000, because I need less heads." — Marc Benioff, Salesforce CEO on The Logan Bartlett Show

A Dallas Fed working paper released the same week showed that the share of employment in high-AI-exposure occupations slipped from 16.4 percent to 15.5 percent since November 2022. The August print is the clearest single macro data point that AI adoption is registering in aggregate labor statistics. The Federal Reserve had penciled a September cut into its June dot plot on inflation grounds; the labor data now provides growth cover for a larger move.

SafetyPolicy

California and Delaware AGs warn OpenAI: safeguards "did not work" after teen suicide, murder-suicide

California OAG / SF Standard / TechCrunch · 2026-09-05 · Source ↗

California AG Rob Bonta and Delaware AG Kathleen Jennings met OpenAI's legal team Friday and delivered a joint letter to Chair Bret Taylor and the board naming the Adam Raine wrongful-death suit and the August Old Greenwich Connecticut murder-suicide of Stein-Erik Soelberg. The two AGs together hold statutory oversight of OpenAI's proposed recapitalization because the nonprofit parent is chartered in Delaware and operates in California.

California Attorney General Rob Bonta and Delaware Attorney General Kathleen Jennings met Friday September 5 with senior members of OpenAI's legal team and delivered a public letter addressed to board Chair Bret Taylor. The letter named two specific tragedies: the Adam Raine wrongful-death case filed in San Francisco Superior Court alleging a young Californian's death by suicide followed prolonged ChatGPT interactions, and the August murder-suicide in Old Greenwich Connecticut in which Stein-Erik Soelberg killed his mother before himself after ChatGPT allegedly reinforced his paranoid delusions.

The two AGs together hold statutory oversight of OpenAI's proposed recapitalization because the nonprofit parent is chartered in Delaware and operates in California — a state-charity-law lever the industry has not fully absorbed. Their meeting escalates the August 25 letter Bonta and 44 other AGs sent to twelve AI firms after reports of sexually inappropriate chatbot interactions with minors. On Tuesday OpenAI announced parental controls, protected-model routing for distressed users, an expert council on well-being and youth safety chaired by David Bickham of Boston Children's Hospital, and expanded clinical review. The Friday meeting was framed as a formal reminder that even absent federal preemption, state charity and consumer-protection powers reach frontier lab governance.

"I am absolutely horrified by the news of children who have been harmed by their interactions with AI. One child harmed is one too many. Whatever safeguards were in place did not work." — joint letter from AGs Rob Bonta and Kathleen Jennings to OpenAI

Taylor responded in a Friday statement that OpenAI is "heartbroken" and reaffirmed safety as the highest priority. The AGs' letter functions as the operative constraint on any recap structure OpenAI ultimately proposes: preservation of the safety mission is a prerequisite for approval. It also puts pressure on parallel legal maneuvers the company was pursuing to convert its nonprofit parent to a benefit corporation.

ApplicationEconomic

OpenAI launches Jobs Platform and Certifications, targets 10M Americans by 2030 with Walmart

CNBC / OpenAI Blog · 2026-09-05 · Source ↗

Fidji Simo, OpenAI's new CEO of Applications, announced an AI-driven matching service that connects employers with candidates OpenAI has certified, plus a paired credential ladder from workplace AI literacy through advanced prompt engineering. Walmart, BCG, John Deere, Accenture and Indeed signed as launch partners; the platform launches by mid-2026 in direct competition with Microsoft-owned LinkedIn.

OpenAI announced the OpenAI Jobs Platform and OpenAI-Certified credential program Friday in a blog post from Fidji Simo, the former Instacart CEO who joined the company in August as CEO of Applications. The platform, launching by mid-2026, will use OpenAI's own models to match candidates with employers on demonstrated AI competencies rather than resume keywords, and it ships with a dedicated track for small local businesses and government agencies alongside enterprise hiring. The framing positions it as a direct competitive threat to Microsoft-owned LinkedIn and to Indeed.

The paired OpenAI Academy offers free certifications spanning basic AI use through prompt engineering and agentic workflow design, integrated as a Study mode inside ChatGPT itself. Target: certify 10 million Americans by 2030. Walmart signed on as launch partner and will offer free OpenAI certification to all US associates starting 2026, committing $1 billion to workforce training through 2026 — the largest private employer in the country (roughly 1.6 million US workers) underwriting the certification standard. Other launch partners include Boston Consulting Group, Accenture, Indeed, John Deere and Texas A&M.

"Jobs will look different, companies will have to adapt, and all of us — from shift workers to CEOs — will have to learn how to work in new ways." — Fidji Simo, CEO of Applications, OpenAI

The strategic read is that OpenAI is converting several hundred million weekly ChatGPT users into a two-sided employment marketplace and defining the credentialing standard for AI-fluent workers before Microsoft, Google or LinkedIn itself defines one. The move also widens the OpenAI-Microsoft strategic rift — OpenAI now competes directly with Microsoft-owned LinkedIn in recruiting software while remaining Microsoft's largest strategic investment target.

Tech

Moonshot AI ships Kimi K2-0905 at 1T parameters with 256K context, Groq serves at 445 tok/s

Simon Willison / Hugging Face · 2026-09-05 · Source ↗

The refresh of July's K2 flagship keeps the 1 trillion parameter mixture-of-experts geometry (384 experts, 32B active per token, 8 selected per token) but doubles the context window from 128K to 256K and posts material gains on coding benchmarks. Weights ship under modified MIT license, immediately available on Fireworks, Groq and OpenRouter.

Moonshot AI shipped Kimi K2-Instruct-0905 on Hugging Face — an incremental refresh of its July 2025 K2 flagship that the community is calling K-2.1. The model keeps K2's mixture-of-experts geometry (1 trillion total parameters across 384 experts, 32 billion activated per token, 8 experts selected per token) but doubles the context window from 128K to 256K tokens and posts material gains on public coding benchmarks and real-world coding-agent tasks. Weights ship under a modified MIT license and are immediately available via Fireworks, Groq, OpenRouter and self-hosted stacks.

Groq is serving the full model at roughly 445 output tokens per second. Simon Willison ran his standard pelican-riding-a-bicycle SVG test through Groq's endpoint and got the completion in just under two seconds. The release matters for three reasons. First, it keeps the open frontier close to the closed frontier — Kimi K2 was already competitive with GPT-5-class systems on agentic coding, and 0905 narrows the gap further while remaining downloadable. Second, 256K context puts K2 in the same operating envelope as Anthropic's Sonnet 4 and Google's Gemini 2.5 for long-horizon coding-agent work. Third, extremely fast inference on Groq's LPUs at open weights is a direct competitive threat to closed-model providers that charge for long-context coding sessions.

"The updated version of K2 improved its performance in coding tasks and increased its context window from 128K tokens to 256K tokens." — Moonshot AI model card

Caveat: the modified MIT license retains a downstream naming and attribution clause that some enterprise legal teams will need to route around, and true reproducibility of the training run is limited by the lack of a released training corpus. The 0905 release lands in the same week Alibaba shipped Qwen3.8-Max, DeepSeek pushed V4-Flash-Vision, and MBZUAI released the K2 Horizon family — a compressed cycle of Chinese open-weights releases that is now the base-rate cadence.

TechApplication

Google releases 308M EmbeddingGemma for on-device multilingual RAG under 200MB

Google Developers Blog · 2026-09-04 · Source ↗

EmbeddingGemma runs in under 200MB of RAM after quantization, produces embeddings in less than 15ms for 256 tokens on EdgeTPU, and covers more than 100 languages with a 2K context window. It ranks first among open multilingual embedding models under 500M on the Massive Text Embedding Benchmark and ships with Matryoshka Representation Learning for 768/512/256/128-dimensional output.

Google released EmbeddingGemma Thursday September 4, a 308-million-parameter open text-embedding model built on the Gemma 3 backbone and engineered specifically for on-device retrieval, classification and clustering. It runs in under 200 MB of RAM after quantization, produces embeddings in less than 15 milliseconds for 256 tokens on EdgeTPU hardware, and covers more than 100 languages with a 2K-token context window. On the Massive Text Embedding Benchmark it ranks first among open multilingual embedding models under 500M parameters and matches models roughly twice its size.

The model uses Matryoshka Representation Learning so a single checkpoint can output 768-, 512-, 256- or 128-dimensional vectors depending on storage and latency budget, and it ships with Quantization-Aware Training so int-quantized versions preserve quality. Launch-day integrations include sentence-transformers, llama.cpp, MLX, Ollama, LiteRT, LMStudio, Weaviate, Cloudflare Workers AI, LlamaIndex, LangChain and transformers.js. Device developers can wire it into a private-by-default RAG stack the day of release.

"EmbeddingGemma is now the highest ranking open multilingual text embedding model under 500M on the Massive Text Embedding Benchmark." — Min Choi and Sahil Dua, Google DeepMind

The strategic significance: RAG has been the last piece of the modern LLM stack that still typically requires a cloud round trip, and EmbeddingGemma pushes it fully into the client. It enables offline semantic search across personal files and email, on-device query classification for mobile agents, and privacy-first chatbot memory. It is Google's most credible open-weights answer to the wave of small Chinese and OpenAI open models pressuring the on-device tier.

TechTrend

Claude produces first end-to-end Lean proof of Fermat's Last Theorem across 11 days

Anthropic Research · 2026-09-04 · Source ↗

Anthropic says Claude produced a 13-million-line Lean proof of Fermat's Last Theorem including at least 29,500 intermediate theorems across 11 continuous days on the Prove2Me multi-agent platform. Andrew Wiles's original 1995 proof was 129 pages of natural-language mathematics; formalizing it in a proof assistant has been a running open target.

Anthropic published a research note Thursday September 4 claiming Claude, operating largely autonomously over 11 days on the Prove2Me platform, produced the first end-to-end computer-checked proof of Fermat's Last Theorem in the Lean proof assistant. The artifact runs to roughly 13 million lines of Lean and includes at least 29,500 intermediate theorems. Anthropic calls it the largest Lean proof ever assembled. Prove2Me, designed by Tianyi Peng and collaborators at Columbia University, maintains a directed acyclic graph of theorem statements and coordinates multiple Claude agents against it, allowing the system to schedule attempts, discharge lemmas, and verify progress mechanically before committing to the shared proof tree.

Andrew Wiles's original 1995 proof was 129 pages of natural-language mathematics and required months of human verification; formalizing it in an interactive theorem prover has been an open target in the mathematical community for a decade. Formal mathematicians reacted with a mix of enthusiasm and skepticism. Kevin Buzzard's Xena Project blog acknowledged Anthropic "beat me to it" on the formalization; others noted the Lean artifact still depends on a large set of imported lemmas from mathlib and questioned how much of the proof is Claude's novel reasoning versus assembly of prior work.

"In 11 days, working largely autonomously, Claude produced the first end-to-end, computer-checked proof of Fermat's Last Theorem. Along the way, it wrote 13 million lines of Lean and proved 29,500 intermediate theorems." — Anthropic Research

Regardless of that debate, the result is a landmark demonstration of long-horizon autonomous agent work — 11 continuous days of tool use, no human intervention on individual proof steps — and it hints at where frontier agent labs are steering: verifiable, formally checkable long-context work products that circumvent the "how do you evaluate agent outputs" problem entirely. Lean-checked artifacts either compile or they don't.

TechSafety

OpenAI paper argues benchmarks — not architecture — cause hallucinations

OpenAI Research · 2026-09-05 · Source ↗

Adam Tauman Kalai, Ofir Nachum, Santosh Vempala and Edwin Zhang argue that even under ideal pre-training on error-free data, models are pressured to hallucinate because generation is intrinsically harder than classification and post-training reinforces guessing because binary-graded benchmarks treat abstentions and wrong answers identically while assigning full credit to lucky guesses.

OpenAI published "Why Language Models Hallucinate" Friday September 5 by Adam Tauman Kalai, Ofir Nachum, Santosh Vempala (Georgia Tech) and Edwin Zhang, arguing hallucinations persist not because of a fundamental limit of language models but because standard training and evaluation procedures actively reward confident guessing over calibrated abstention. The paper proves that even under ideal pre-training on error-free data, models are pressured to hallucinate because generation is intrinsically harder than classification — the generative error rate is at least twice the classification (Is-It-Valid) error rate.

Post-training reinforces guessing because the dominant binary-graded benchmarks (MMLU-Pro, GPQA, SWE-bench, HLE) treat abstentions and wrong answers identically while assigning full credit to lucky guesses. The proposed fix: modify benchmark scoring rubrics to explicitly reward "I don't know" whenever a model's confidence sits below a stated threshold, mirroring negative marking used on the SAT and NMSQT. GPT-5 in reasoning mode is shown to hallucinate on 26 percent of open-ended questions in an internal SimpleQA-style eval while GPT-4o hallucinates on 52 percent.

"Hallucinations persist due to the way most evaluations are graded — language models are optimized to be good test-takers, and guessing when uncertain improves test performance." — OpenAI blog post

The paper drops as OpenAI's Preparedness Framework classifies advanced hallucination-triggered harms as a distinct risk vector, and it is co-timed with California SB 53 (frontier-AI transparency) advancing through Sacramento. Externally, the paper is read as OpenAI trying to move the evaluation industry — Artificial Analysis, LMSYS Arena, HELM — toward calibrated scoring rather than pure accuracy leaderboards.

EconomicTech

Baseten raises $150M Series D at $2.15B as inference platform revenue jumps 10×

Fortune / Baseten · 2026-09-05 · Source ↗

The AI inference platform behind Abridge, OpenEvidence, Clay, Patreon and Writer closed a $150 million Series D Friday at a $2.15 billion post-money led by Bond, with CapitalG (Alphabet's growth-stage fund), Premji, and Scribble family office joining. The round nearly triples Baseten's valuation from its $75M Series C six months ago; total capital now exceeds $285 million.

Baseten, the AI inference platform behind fast-growing production apps like Abridge, OpenEvidence, Clay, Patreon and Writer, closed a $150 million Series D on Friday September 5 at a $2.15 billion post-money valuation. Bond led the round; new investors include CapitalG (Alphabet's growth-stage fund), Premji Invest, and the Weil family office Scribble. Existing backers Conviction, 01a, IVP, Spark and Greylock also participated. The round nearly triples Baseten's valuation from six months ago when it raised its $75M Series C, and brings total capital raised to more than $285 million.

Baseten disclosed that revenue grew more than 10x over the past twelve months as customers increasingly separated inference workloads from training clouds to chase lower latency and cost. The round lands the same week Anthropic closed $13 billion at $183 billion and Sierra raised $350 million at $10 billion-plus, extending an unbroken run of unicorn-scale AI infrastructure financings and confirming inference — not training — as where the next capex layer is being underwritten.

"Inference will be one of the biggest markets, and Baseten has the leading product." — Sarah Guo, Conviction founder

Baseten sits in a competitive field including Together AI, Fireworks AI, Modal, Anyscale and Replicate. Its focus on latency-critical enterprise apps — particularly healthcare with Abridge and OpenEvidence — differentiates it from broader model-hosting rivals. The Bond-led round marks Jay Simons' second AI infra bet after his Anthropic board seat; Simons called Baseten "a critical service that needs to be absolutely bullet-proof."

ApplicationTech

Tesla Robotaxi app crosses 2M downloads on launch day, outpaces Uber and Waymo

Tesla Oracle / Ark Invest · 2026-09-05 · Source ↗

Tesla's standalone Robotaxi app, split from the main Tesla vehicle app the day before, crossed two million iOS downloads within its first 24 hours of public availability, taking the #1 free spot in the App Store's Travel category with a 4.9-star rating across 355 reviews. Ark Invest said the day-one performance outpaced any single 30-day rolling install month from either Uber or Waymo.

Tesla's standalone Robotaxi app, split off from the main Tesla vehicle app the day before, crossed two million iOS downloads within its first 24 hours of public availability on the App Store. Ark Invest analysis said the number outpaced any single 30-day rolling install performance from either Uber or Waymo on their best months by a substantial margin. The app rocketed to the #1 free spot in the App Store's Travel category within hours and holds a 4.9-star rating across 355 initial reviews.

Tesla shipped two versions in a single day: 25.8.10 released September 4 as a preview build, then 25.9.0 pushed September 5 to enable ride-hailing through the standalone client. The rollout coincides with Elon Musk's decision to send driverless Cybercabs — the two-seat vehicle Tesla unveiled with no steering wheel and no pedals — onto Austin streets September 3, then formally opening rides to the public app the following day at 2 PM Central. Cybercabs join the existing Model Y-based Robotaxi fleet operating in Austin (safety personnel in the passenger seat) and the San Francisco Bay Area (safety operator in the driver's seat per California regulations).

"The safety driver is just there for the first few months to be extra safe. Should be no safety driver by end of year." — Elon Musk, on removing safety operators from Tesla Robotaxi

Tesla shares rose $9.30 (2.76 percent) the day of the app milestone, closing 24 hours in which Tesla materially closed the perception gap with Waymo despite continuing federal investigations of its Full Self-Driving stack. An Android version is coming, Tesla told users in the app-store listing. Ark's read: this is the fastest-adopted new mobility app since Uber's original 2010 launch, and the first time a robotaxi-native app has crossed a million-plus downloads in a single day.

ApplicationTech

Samsung debuts Vision AI Companion across 2025 TV and monitor lineup at IFA Berlin

Samsung Global Newsroom · 2026-09-05 · Source ↗

The conversational AI layer fuses on-device Live Translate, AI Upscaling and Voice ID user recognition with cloud agents handling Q&A, content discovery and productivity through a smarter Bixby, across Samsung's 2025 Neo QLED, OLED, QLED, Micro RGB TVs, smart monitors and Movingstyle series with a promised software rollout to 2023 and 2024 sets. Samsung committed seven years of free software upgrades.

Samsung Electronics used the IFA Berlin opening Friday September 5 to unveil Vision AI Companion, a conversational AI layer that turns Samsung's 2025 TVs and monitors into always-listening assistants. The system fuses on-device features — Live Translate, AI Upscaling, Voice ID user recognition — with cloud agents handling conversational Q&A, content discovery, follow-up recommendations, and productivity tasks routed through a smarter Bixby.

Ask "what movie should we watch tonight," and Vision AI Companion returns curated suggestions with poster art, showtimes and streaming availability rendered directly on the display; ask a follow-up like "something shorter, no violence" and it refines without repeating context. Voice ID lets the assistant recognize which household member is speaking and tailor responses accordingly. The platform is available across Samsung's 2025 Neo QLED, OLED, QLED, Micro RGB TVs, smart monitors and Movingstyle series, with 2023 and 2024 smart TVs receiving the feature through a future Tizen OS update.

"Vision AI Companion responds to context and follow-up interactions, displaying relevant information directly on the screen, with content spanning movies, dramas, art, food, sports and travel in conversational form." — Samsung Global Newsroom

Samsung committed seven years of free software upgrades through the One UI Tizen platform including AI feature additions and Samsung Knox security patches. Vision AI Companion begins rolling out as a software update in late September starting in Korea, North America and select European markets, with additional regions to follow. Samsung also announced the 2025 Bespoke AI Steam robot vacuum (RGB camera and IR LED sensors detecting transparent liquids), an AI hybrid refrigerator, and an all-in-one washer-dryer under the same "AI Home: Future Living, Now" IFA theme.

ApplicationTech

Neura Robotics puts 4NE1 humanoid on IFA Berlin floor sorting laundry live

BGR / Neura Robotics · 2026-09-05 · Source ↗

German cognitive robotics company Neura Robotics opened IFA with live consumer demos of its six-foot 4NE1 humanoid sorting whites from colors into laundry baskets, alongside the MiPA wheeled mobile manipulation platform tidying stuffed animals. Head of engineering Dr.-Ing. Arne Nordmann said consumer entry is within "a couple of years" and pledged no military sales.

German cognitive robotics company Neura Robotics opened IFA Berlin on September 5 with live consumer demonstrations of its 4NE1 humanoid ("for anyone") — the six-foot, roughly 176-pound bipedal robot Germany has positioned as its answer to Tesla's Optimus and Figure 03 — sorting white clothes from colored ones into multiple laundry baskets in front of trade-show attendees at Hall 25 Booth H25-181. Alongside 4NE1, Neura showed MiPA (My intelligent Personal Assistant), a wheeled mobile manipulation platform, tidying stuffed animals in a mocked-up child's bedroom.

Dr.-Ing. Arne Nordmann, Neura's head of engineering, told BGR the company builds for "dirty, dull, or dangerous" work — cleaning warehouses, monotonous factory-line jobs, hazardous inspection — and expects consumer entry within a couple of years as physical AI capabilities catch up to the hardware. Neura's differentiator is a cognitive intelligence stack that lets the robot perceive and react to unstructured environments without safety cages, adapt grasps it wasn't explicitly trained on, and swap dual batteries for continuous 24/7 operation. The robot is rated to lift 100 kg (220 lb) thanks to Neura's in-house joint actuators.

"We see them trying out new grasps that we didn't explicitly train. That's the magic of physical AI. We serve humanity, and that means no involvement in military activities." — Dr.-Ing. Arne Nordmann, Head of Engineering, Neura Robotics

The IFA showing joins Beijing Galbot (running live retrieval demos at IFA Next Booth 160, already handling 5,000 SKUs in a Beijing Haidian pharmacy at 99.5 percent pick success) and a scheduled "Robots on the Runway" catwalk at 1 PM on the Creator Stage with humanoids from Unitree, Agibot, DEEP Robotics, EngineAI and Dobot — IFA's clearest signal yet that consumer-adjacent humanoids are moving from R&D lab to trade-show demo floor.

ApplicationTech

Lenovo unveils widest ThinkPad AI PC refresh — P16 Gen 3 up to 192GB and RTX PRO 5000 Blackwell

Lenovo StoryHub · 2026-09-05 · Source ↗

The ThinkPad P16 Gen 3 mobile workstation ships Intel Core Ultra 200HX (up to 24 cores, 5.5 GHz) with integrated NPU, up to NVIDIA RTX PRO 5000 Blackwell with GDDR7, up to 192 GB DDR5-5600, up to 12 TB PCIe Gen5 NVMe and a 99.9 Whr battery — a spec sheet targeted at engineers running local LLM inference and CAD without cloud round-trips. Lenovo currently holds ~30% of AI PC market and expects 50% within 18 months.

Lenovo used its Innovation World 2025 keynote in Berlin on September 5 — timed to coincide with IFA opening day — to unveil the widest ThinkPad AI PC refresh in the brand's history. The centerpiece is the ThinkPad P16 Gen 3 mobile workstation with Intel Core Ultra 200HX series processors (up to 24 cores, 5.5 GHz), integrated NPU for on-device AI acceleration, up to NVIDIA RTX PRO 5000 Blackwell discrete graphics with GDDR7, up to 192 GB DDR5-5600 memory, up to 12 TB of PCIe Gen5 NVMe SSD, and a 99.9 Whr customer-replaceable battery — a spec sheet targeted at engineers running local LLM inference and CAD workloads without cloud round-trips.

Availability: P16v Gen 3, P16s i Gen 4 and P14s i Gen 6 land in September starting at €1,569; P16 Gen 3 and P1 Gen 8 arrive in October from €2,899. Lenovo paired the workstation refresh with a Glacier White TrackPoint-less ThinkPad X9 Aura, three new ThinkPad Thunderbolt 5 Smart Docks (from €229), the ThinkVision P40WD-40 ultrawide display (November, €1,499), the Legion Go Gen 2 handheld with detachable controllers on AMD Ryzen Z2 Extreme silicon and an 8.8-inch OLED, and a series of foldable phones, AR glasses and laptop concept prototypes.

"With our expanding AI PC portfolio, forward-thinking concepts, and enterprise-grade support, we're enabling organizations to confidently embrace the AI era on their terms." — Eric Yu, SVP and GM, Lenovo Commercial Product Center

Lenovo currently holds roughly 30 percent of the AI PC market and told analysts at the event it expects to hit 50 percent within 18 months. The message is that AI compute in the enterprise is moving to the edge device, not just the data center.

Policy

Apple sued by novelists Hendrix and Roberson over Books3 training — 49th AI copyright case

Publishers Weekly / CNBC · 2026-09-05 · Source ↗

Horror novelist Grady Hendrix and fantasy author Jennifer Roberson filed a putative class action in the Northern District of California Friday alleging Apple used the Books3 pirated corpus and its Applebot crawler to scrape shadow libraries to train the OpenELM family powering Apple Intelligence. The suit is the 49th US AI copyright case and the 18th filed by book authors.

Horror novelist Grady Hendrix (Witchcraft for Wayward Girls) and fantasy author Jennifer Roberson filed a putative class action in the US District Court for the Northern District of California Friday September 5, alleging Apple used the notorious Books3 pirated corpus and its proprietary Applebot web crawler to scrape books from shadow libraries to train the OpenELM family of models powering Apple Intelligence. The complaint — represented by Susman Godfrey, the same firm that struck the $1.5 billion Anthropic settlement hours earlier the same day — accuses Apple of copying protected works "without consent, credit, or compensation" and calls out the internal M4-hosted OpenELM series unveiled at WWDC24.

The filing is the 49th US copyright lawsuit against an AI company since the New York Times v. OpenAI complaint in December 2023, and the 18th brought by book authors. The concentration reflects how the class-action bar has now systematically mapped every major frontier lab against the same LibGen/PiLiMi source-of-truth. Plaintiffs seek statutory damages, disgorgement and destruction of infringing models — the remedy Judge William Alsup approved in the Anthropic term sheet, effectively creating a template settlement demand at $3,000 per book against 2,500 to 500,000 titles.

"Apple has not attempted to pay these authors for their contributions to this potentially lucrative venture." — Hendrix and Roberson complaint

Apple joins OpenAI, Anthropic, Meta, Microsoft, Cohere, Perplexity, Nvidia, Databricks, Suno, Udio and Stability as named defendants. The Hendrix and Roberson filing landed the same Friday that Judge Alsup was reviewing the Anthropic $1.5 billion term sheet in the same courthouse.

PolicySafety

Pennsylvania Act 35 criminalizes forged AI digital likenesses statewide

Crowell & Moring / PA Senate · 2026-09-05 · Source ↗

Pennsylvania Act 35 of 2025 took effect Friday, 60 days after Governor Josh Shapiro's July 7 signature, criminalizing the creation or distribution of "forged digital likenesses" — synthetic visual images or audio of a real, identifiable person made without consent and passed off as genuine. Violations are first-degree misdemeanors ($1,500-$10,000 fines, up to 5 years); escalating to third-degree felonies (up to $15,000 and 7 years) when intent is fraud, coercion or theft.

Pennsylvania Act 35 of 2025 took effect Friday September 5, 60 days after Governor Josh Shapiro's July 7 signature, criminalizing the creation or distribution of "forged digital likenesses" — synthetic visual images or audio recordings of a real, identifiable person made without consent and passed off as genuine, with intent to defraud or injure. Sponsored by state Sen. Tracy Pennycuick (R-Montgomery), the statute amends the Pennsylvania forgery framework (18 Pa. C.S. § 4101) to add a new category covering both AI-generated video and voice cloning outside the robocall context.

Pennsylvania becomes one of a small handful of US states to criminalize AI voice cloning at that scope. Violations are graded as first-degree misdemeanors carrying $1,500-$10,000 fines and up to five years imprisonment, escalating to third-degree felonies (up to $15,000 fines and seven years) when done to defraud, coerce or commit theft. Washington state's Deceptive Synthetic Media Act (HB 1205) took effect the same July date on a parallel track.

"Deepfakes are a growing threat, and Pennsylvanians deserve strong protections against this kind of digital fraud." — Sen. Tracy Pennycuick (R-Montgomery), bill sponsor

Act 35 arrives against a widening federal patchwork: the TAKE IT DOWN Act became law in May, Tennessee's ELVIS Act protecting voice actors is a year old, and the No FAKES Act cleared Senate Judiciary in July. Pennsylvania's addition puts roughly two-thirds of US states under some form of criminal or civil deepfake statute as the 2026 election cycle approaches.

PolicyEconomic

WSJ: US dangled Nvidia Blackwell access to Armenia to broker Azerbaijan peace deal

Wall Street Journal · 2026-09-05 · Source ↗

The Journal reports US negotiators explicitly linked Armenia's ability to import advanced Nvidia Blackwell GPUs to signing the August peace declaration, via a Firebird export-license approval for a 100 MW AI supercomputing site in Armenia using roughly 6,000 Nvidia GPUs. The arrangement formalizes an emerging US doctrine of "chip diplomacy."

The Wall Street Journal reported Friday September 5 that US negotiators, during the run-up to the White House-brokered Armenia-Azerbaijan peace declaration in August 2025, explicitly linked Armenia's ability to import advanced Nvidia Blackwell GPUs to its willingness to sign. The specific instrument was an export-license approval for US startup Firebird, which is building a planned 100 MW AI supercomputing site in Armenia and had applied to buy roughly 6,000 Nvidia GPUs for the facility. Chip export approval reportedly followed the signing.

The Journal frames the arrangement as an emerging US doctrine of "chip diplomacy," in which access to top-bin AI accelerators is treated as a foreign-policy carrot on par with military aid or trade concessions. Prime Minister Nikol Pashinyan's spokesman confirmed the AI datacenter play was part of the diplomatic package. The report matters for the AI industry because it formalizes what industry watchers had suspected — Blackwell allocations are now geopolitical currency — and previews a regime where individual export licenses become deal-linked rather than purely technology- or end-use-based.

"This was chip diplomacy — access to advanced US technology used as leverage in a peace deal." — Wall Street Journal characterization of the negotiation

The precedent will not be lost on other emerging economies. The fastest way to secure frontier compute may now run through the National Security Council rather than through Nvidia's sales team, and datacenter siting decisions are increasingly a diplomatic instrument. The Armenia case joins a growing set of chip-linked deals including the Saudi HUMAIN commitment and UAE G42-Microsoft arrangement.

TechApplication

Warp launches Warp Code, an agent-native IDE layer atop its terminal

Warp · 2026-09-04 · Source ↗

Warp Code adds a built-in file editor with real-time diff view, code-review panel with inline approve/reject, and a WARP.md project-configuration file cross-compatible with AGENTS.md, CLAUDE.md and Cursor rules. Warp claims #1 on Terminal-Bench at 52% and top-three on SWE-bench Verified at 75.8% driving GPT-5, with users producing 150M+ lines of code per week and a 97%+ diff acceptance rate.

Warp, the Rust-built agentic terminal founded by ex-Google engineer Zach Lloyd, launched Warp Code on Thursday September 4 — a set of features that turn Warp from an AI terminal into a full prompt-to-production coding environment. New capabilities include a built-in file editor with real-time diff view, a code-review panel where developers approve or reject agent-generated changes inline, a WARP.md project-configuration file that is cross-compatible with AGENTS.md, CLAUDE.md and Cursor rules, and Agent Profiles that scope which models, tools and permissions each agent gets.

Warp claims its coding agent sits at #1 on Terminal-Bench (52 percent) and top-three on SWE-bench Verified (75.8 percent when driving GPT-5) as of September 2. Users are producing over 150 million lines of code per week inside Warp with a 97-percent-plus diff-acceptance rate. The launch positions Warp as a direct competitor to Cursor and Anthropic's Claude Code — the three now define the "agentic development environment" (ADE) category, with Warp taking the terminal-first angle, Cursor the IDE-first angle, and Claude Code the pure-CLI angle.

"Too often agents write code that almost works, but doesn't. Warp Code is the fastest way from prompt to production." — Zach Lloyd, CEO of Warp

Warp's stack is model-agnostic, running on GPT-5 and Claude 4 Sonnet interchangeably, which insulates it from any one lab's pricing or policy changes and makes Terminal-Bench a useful cross-model benchmark. The company's decision to publish a WARP.md file spec that reads and writes to AGENTS.md is a bet that a shared project-configuration standard emerges across the ADE category.

TechTrend

Chris Lattner argues ML needs a new programming language, HN top of front page

Hacker News / Modular · 2026-09-05 · Source ↗

The LLVM, Clang, Swift and MLIR creator argues Python's grip on ML is now a limiting factor: CPython's interpreter, C-extension calling conventions and CUDA glue make it impossible to write portable high-performance ML code that runs across Nvidia, AMD, Apple Silicon and custom ASICs without vendor-specific rewrites. Mojo is his answer — a Python superset with MLIR compilation, ownership and first-class GPU primitives.

An interview with Chris Lattner titled "ML needs a new programming language" hit the Hacker News front page Friday September 5, pulling 311+ points and 278 comments over the day. Lattner — creator of LLVM, Clang, Swift and MLIR, and co-founder and CEO of Modular AI, which builds the Mojo language and MAX inference engine — argues that Python's grip on ML is now a limiting factor rather than an enabler. The CPython interpreter, the C-extension calling conventions, and the CUDA-specific glue code around PyTorch collectively make it impossible to write portable, high-performance ML code that runs across Nvidia, AMD, Apple Silicon and custom ASICs without rewriting hot paths in vendor-specific C++ or CUDA.

Mojo is his proposed answer: a Python superset with strong ownership, static compilation via MLIR, first-class GPU programming primitives, and a hardware-abstraction layer that targets multiple accelerator backends from one source. The HN discussion focused on whether the ecosystem inertia around PyTorch and CUDA is now too large to displace, and on whether Modular's commercial licensing terms will slow adoption relative to a fully open BSD or MIT release.

"Mojo is designed to combine the simplicity of Python with the performance of C and the safety of Rust. We need a language that lets you target any accelerator from a single source of truth." — Chris Lattner, CEO of Modular

The story is significant because Mojo has now moved from "interesting research project" to "actively pitched CUDA replacement," and Lattner is publicly framing hardware fragmentation — AMD MI300/MI400, Google TPU, AWS Trainium, Meta MTIA, Broadcom-custom XPUs — as the wedge that gives a new language a real shot.

TechApplication

Simon Willison christens GPT-5 Thinking with browsing as the "Research Goblin"

Simon Willison's Weblog · 2026-09-06 · Source ↗

Willison argues GPT-5 Thinking with browsing has crossed the threshold from "chatbot with search" into a genuinely useful open-ended research tool. It executes a search, reads results, spawns follow-up searches, cross-checks against PDFs and specialized databases, and continues reasoning across all of that within a single thinking session with the chain-of-thought exposed for methodology audit.

On Saturday September 6, Simon Willison published an essay arguing that GPT-5 Thinking coupled with ChatGPT's browsing tool has quietly crossed the threshold from "chatbot with search" into a genuinely useful open-ended research tool that he nicknames the "Research Goblin." The mechanism is what Anthropic separately calls "interleaved thinking": the model can execute a search, read the results, spawn follow-up searches, cross-check against PDFs and specialized databases, and continue reasoning across all of that within a single thinking session, with the chain-of-thought exposed so the user can audit methodology.

Willison walks through nine worked examples: dating Heathrow's travelator replacement (2014-2018), identifying a building from a phone photo, tracing the 1911 Britannica-to-Wikipedia seeding project, and pulling planning documents and PDFs for Exeter Quay research. Research Goblin failed only on data-visualization tasks and occasionally overreached on follow-up suggestions; the mobile UX (voice input, hands-free) is where the tool genuinely displaces manual Google and Wikipedia workflows.

"I can assign a task to it, no matter how trivial or complex, and it will do an often unreasonable amount of work to search the internet. It's very industrious, not quite human and not entirely trustworthy. You have to be able to outwit it if you want to keep it gainfully employed." — Simon Willison

In an aside relevant to the Anthropic settlement, Willison had Research Goblin scan for evidence of other labs doing large-scale book scanning and reports it "was unable to find any news stories or other evidence that any labs other than Anthropic are engaged in large scale book scanning." Significance: this is one of the sharpest working-developer descriptions of what "reasoning + tools" now does in production, and the Research Goblin framing was widely adopted in the days after publication as shorthand for the new capability tier.

TechApplication

Redis to acquire Decodable, launches LangCache semantic cache for LLM apps

Redis / GlobeNewswire · 2026-09-04 · Source ↗

At Redis Released 2025 Thursday, CEO Rowan Trollope announced Redis will acquire Decodable, an Apache Flink-based streaming platform, to expand "Redis for AI" into the memory-and-context layer for agentic systems. LangCache, a managed semantic cache for LLM applications, entered public preview claiming up to 70% reduction in latency and token cost for repeated or near-repeated prompts.

At Redis Released 2025 on Thursday September 4, CEO Rowan Trollope announced Redis will acquire Decodable, a real-time streaming data platform built on Apache Flink, to expand "Redis for AI" — the company's push to become the memory-and-context layer for agentic systems. Alongside the M&A, Redis opened a public preview of LangCache, a fully-managed semantic cache for LLM applications that Redis says cuts latency and token cost by up to 70 percent for repeated or near-repeated prompts.

The Decodable deal folds streaming ingestion and change-data-capture pipelines into the Redis stack, so an agent framework built on Redis can subscribe to fresh operational data (Kafka topics, database CDC streams, event buses) without a separate ETL layer. The combined pitch is that context and memory for AI agents — vector recall, semantic caching, short-term working memory, long-term episodic memory, and now real-time event streams — can all live inside the Redis boundary, competing directly with LangChain/LangGraph-plus-Kafka-plus-pgvector stacks.

"Real-time data is what turns generic AI models into intelligent, contextual applications. With Decodable and LangCache, Redis is delivering the context and memory that intelligent agents depend on." — Rowan Trollope, CEO of Redis

This is Redis's most explicit repositioning to date from "the fastest in-memory data store" to "AI-agent infrastructure," and it lands in a market where every vector database, streaming platform and RAG framework is racing to claim the same layer.

Tech

Cloudflare's Kenton Varda reframes Durable Objects as "StatefulWorker"

Simon Willison / Cloudflare · 2026-09-05 · Source ↗

The Cloudflare Workers lead engineer proposes renaming Durable Objects to "StatefulWorker" to make the mental model obvious: a named Worker instance that holds persistent state (SQLite DB, container filesystem, in-memory state) that Cloudflare's runtime routes messages to by name across the global fleet. Durable Objects now underpin per-user agent sessions, per-conversation memory and per-tenant MCP sandboxes.

On Friday September 5, Simon Willison surfaced comments from Cloudflare's Kenton Varda — the lead engineer behind Workers and Durable Objects — arguing that Cloudflare's Durable Objects primitive is not conceptually complex, only badly named. Varda proposes renaming them "StatefulWorker" to make the mental model obvious: a Durable Object is a named Worker instance that holds persistent state (a SQLite database, container filesystem, or in-memory state), that Cloudflare's runtime routes messages to by name across the global fleet.

The reframing matters because Durable Objects are one of the most-used primitives in production agent infrastructure right now. They underpin per-user agent sessions, per-conversation memory, per-tenant sandboxes for MCP servers, and long-running WebSocket connections for real-time agent UIs. Several agent frameworks — including Cloudflare's own Agents SDK, and third-party stacks like Superduper — model an "agent session" as one Durable Object per user, giving each conversation its own consistent SQLite database and stateful actor.

"I think Durable Objects would be a lot less scary if we just called them StatefulWorker. A Durable Object is just a worker with persistent state." — Kenton Varda, Cloudflare

Varda's post reframes them as the AI-native cloud primitive for stateful, single-writer, globally-addressable agents — a positioning that puts Cloudflare's Workers platform directly opposite AWS Lambda plus DynamoDB and Vercel Fluid for the agent-runtime tier.

Application

Microsoft flips Copilot in Power BI to default-on for tenants with Copilot enabled

Microsoft 365 Copilot Blog · 2026-09-05 · Source ↗

Microsoft is turning the standalone Copilot experience in Power BI on by default for every tenant that already has Microsoft 365 Copilot enabled. The full-screen conversational interface lets business users ask multi-turn questions across all their semantic models, and Copilot retrieves data, generates visuals, and returns cited answers. Admins retain the ability to turn it back off.

Microsoft announced Friday September 5 that the standalone Copilot experience in Power BI is being turned on by default for every tenant that already has Microsoft 365 Copilot enabled. The standalone Copilot in Power BI is a full-screen conversational interface (as opposed to the in-report Copilot pane) where a business user can ask multi-turn questions across all their semantic models, and Copilot retrieves data, generates visuals, and returns answers with citations back to the underlying dataset.

Making it default-on is a material posture change: Microsoft is treating Copilot as the primary interaction surface for BI, not a bolt-on. The rollout is gated by existing per-workspace Copilot governance and the tenant-level Copilot admin toggle, and admins retain the ability to turn it back off.

"Copilot in Power BI standalone experience is on by default for all organizations that have Copilot enabled. Users can conversationally ask questions about their data and reports, and Copilot will retrieve data, generate visuals, and provide answers." — Microsoft 365 Copilot September changelog

For enterprise data teams, this flips the "who has access to Copilot in Power BI" conversation from opt-in to opt-out and forces IT organizations to formalize governance (row-level security, semantic model preparation, prompt logging) more quickly than they otherwise would have. It is one of the clearest signals Microsoft has sent that Copilot has, in the company's view, moved past the "measure adoption" phase and into the "default UX" phase for its enterprise data stack.

TechApplication

Anthropic broadens Claude for Chrome to Pro, Max, Team and Enterprise plans

Anthropic · 2026-09-05 · Source ↗

Until Friday the Claude for Chrome extension was a research preview limited to Max subscribers; Anthropic broadened it to every paid Claude plan. Claude can now read and interact with the current browser tab: inspect DOM state, click buttons, fill forms, extract structured data, debug running JavaScript and drive multi-step workflows. Team and Enterprise admins get org-wide toggles and URL allowlists.

Anthropic on Friday September 5 broadened the Claude for Chrome extension — until then a research preview limited to Max subscribers — to every paid Claude plan, including Pro, Max, Team and Enterprise. The extension lets Claude read and interact with the current browser tab: it can inspect DOM state, click buttons, fill forms, extract structured data, debug the running JavaScript, and drive multi-step workflows across the pages the user has open.

For Team and Enterprise, admins get org-wide enable/disable toggles plus URL allowlists and blocklists so security teams can restrict the extension's reach to a defined set of internal apps. The launch matters as one of the first widely deployed computer-use surfaces from a frontier lab where the model actually drives an end-user's browser, rather than a remote sandboxed browser controlled by the lab.

"Claude in Chrome allows Claude to read, click, and navigate websites alongside you. On Team and Enterprise, admins turn the extension on or off org-wide and set site allowlists and blocklists." — Anthropic product page and documentation

It puts Anthropic in direct competition with OpenAI's Operator, Google Project Mariner and Perplexity's Comet browser for control of the AI-in-the-browser tier, and it makes Chrome — not a bespoke app — the front-end for Claude's agentic capabilities. Caveat: the extension inherits every prompt-injection vulnerability that comes with letting an LLM read arbitrary web pages; Anthropic's system-card update pairs the launch with explicit guidance on scoping site access and running the extension against only trusted domains.