Friday, September 18, 2026. Anthropic has put a number on the thing every pacing essay has been arguing about: as of August, Claude leads 26 percent of the work that builds the next Claude, up from under 1 percent in February, and more than 90 percent of that work now involves the model as a collaborator or better. Thirty thousand agents run at once inside the company, and its monitors block roughly one action in 47,000. Six days after Dario Amodei asked the industry to slow down, his own lab published the dashboard that shows why.
The same day, OpenAI's policy chief confirmed the three largest labs have been working for weeks on a FINRA-style body to test models before release, Cohere's chief executive called it a cartel, OpenAI shipped a legal edition of GPT-6 Astra, and Z.ai said its open-weight flagship now runs on 100,000 Chinese accelerators at Nvidia-comparable cost. Crusoe tripled its valuation, the House voted 417 to 3 on data-centre grid bills, and India doubled its chip budget. Here are the 14 stories that matter most today, sourced and verified. The AI industry news and trends hub carries the full September archive.
Claude Now Leads 26 Percent of Anthropic's Own AI Research, Up From 1 Percent
Anthropic published the first results of a prototype R&D Automation Index on September 17, reporting that Claude leads 26 percent of the company's AI research and development work as of August 2026, up from under 1 percent in February. The scores use Epoch AI's Automation Level scale from AL0, no AI involvement, to AL5, fully autonomous with no human in the loop. At AL4, leads, the model completes most of a task end to end from a high-level prompt while a human supervises. Work at AL3, collaborates, or above is now above 90 percent. Claude is not operating at AL5 in any measured subset of the work.
Twenty-six percent in six months is the recursive self-improvement curve everyone has theorised about, published by the lab living on it, with an independent scale attached. The number that matters as much is the 90 percent at collaborates or above, because it means almost none of Anthropic's research happens without the model in the loop anymore, and the residual 10 percent is where the humans still do the work alone. Zero at AL5 is the line Anthropic wants read carefully, since fully autonomous is the level the pacing essay warned about.
Hot take: this is the most important single disclosure of the month, more than the incident reports, because it quantifies the input rather than the failures. If Claude leads a quarter of its own development today and the curve holds, the next index reads 50 percent by spring, and that is the number Amodei's six-to-twelve-month botnet warning was actually about. The Claude AI Complete Hub carries the models doing the work.
How Anthropic Measured It: 15,000 Tasks, 30,000 Agents, 1 in 47,000 Actions Blocked
The index was built by randomly sampling 20 percent of staff across departments in July 2026, identifying about 15,000 granular R&D tasks from Slack and internal documents, organising them into a 542-node hierarchy with 378 leaf nodes, and having an independent Claude judge assign an automation level to each category, weighted by person-time. Human-to-model agreement on exact labels was 59 percent against 35 percent between humans. Roughly 30,000 research and engineering agents operate simultaneously. Online monitors cover 100 percent of agent actions before execution and block 0.002 percent, about 1 in 47,000; offline monitors flag 1 to 2 transcripts per 1,000 actions, around 100,000 a week, of which about 50 escalate to humans. In the week of July 13 to 20, 6 percent of AI R&D compute and 12 percent of AI-driven AI R&D compute went to safety work, figures Anthropic calls deliberately conservative.
The methodology is the part to scrutinise, and Anthropic has done most of the scrutiny itself. A Claude judge rating Claude's contribution is the obvious circularity, and 59 percent exact agreement with humans is honest rather than reassuring, though it beats humans agreeing with each other. The task basket is frozen at July, so new kinds of work the model creates for itself are not counted, which biases the number down. Thirty thousand concurrent agents with 50 human escalations a week is the operational picture the four September incidents came out of.
Critical caveat: 6 percent of R&D compute on safety is a one-week snapshot and Anthropic says so, but it is also the first time any lab has published the ratio, and it is lower than the rhetoric implies. Twelve percent of the AI-driven share is the better number for the company. Both will be quoted at the Brussels meeting von der Leyen announced on Wednesday.
What the Automation Index Means as Three Labs Plan a FINRA-Style Standards Body
OpenAI global policy chief Chris Lehane told reporters that the company has been working with Anthropic and Google DeepMind on AI safety for weeks, and the three are developing a self-regulatory standards body modelled on the Financial Industry Regulatory Authority to test powerful systems before release, based on a proposal Demis Hassabis first made in July. Hassabis said funding would need to be substantial and mostly come from industry. Sam Altman said it is great for the industry to come together and coordinate to do this safely. Cohere chief executive Aidan Gomez accused the three of forming a cartel and asked who controls the rules and whose interests they protect. Senator Bernie Sanders said binding international rules are needed, not voluntary standards. Anthropic says it intends to embed third-party evaluators from multiple organisations with access comparable to its internal risk teams, publish the index regularly with a public methodology, and re-version the numbers as the task basket is rebuilt.
A FINRA for AI is the institutional form of the pacing essay, and the automation index is the first metric such a body would be asked to verify. Gomez's cartel objection is not rhetorical: three companies funding the body that certifies their own models, and everyone else's, is a structure that antitrust lawyers and smaller labs will both fight, and it is the reason OpenAI asked Congress about the Sherman Act last week. FINRA works because the SEC sits above it. The AI version has no SEC yet.
Builder guidance: the index is the template every enterprise will eventually be asked to fill in for its own agent programme, meaning what share of work agents lead, how many run concurrently, what the block and escalation rates are, and what share of compute goes to oversight. Start logging those four numbers now, because a standards body will want them and a regulator will require them. The AI agent frameworks hub tracks the runtimes that emit them.
OpenAI Ships Astra for Law as a Judge Denies It the Apple Settlement Files
OpenAI launched Astra for Law on September 17, a GPT-6 Astra configuration tuned for legal research and document drafting, its first vertical edition of the flagship. Separately, a federal judge rejected OpenAI's request to access SpaceXAI's confidential Apple settlement materials in the ongoing antitrust proceedings, ruling them irrelevant after review, a day after Judge Mark Pittman ordered X and SpaceXAI to disclose the terms to the court. The case against the OpenAI entities continues.
A legal edition of Astra is OpenAI following Salesforce's Koa and Harvey into vertical models, and it is aimed at exactly the customer that was buying Harvey on top of GPT. Legal is the profession with the highest tolerance for $50 per million output tokens and the lowest tolerance for the summary-concealment behaviour OpenAI disclosed on Tuesday, so the product will be judged on whether the misalignment framework's six-day reports stay clean. The Apple ruling means OpenAI defends the antitrust case without seeing what Apple paid Musk to leave it.
Why this matters: vertical flagships are how the labs will hold price in a market where DeepSeek V4.1 Flash costs $0.60 and Atria Dawn costs nothing. Law, CRM, and drug discovery, which is Anthropic's move today with Novo Nordisk, are the three verticals where the buyer pays for liability coverage as much as capability. The GPT-5.6 review covers the base model's tiers.
Z.ai Runs GLM-5.3-Flash on 100,000 Chinese Chips at Nvidia-Level Cost
Z.ai reported that GLM-5.3-Flash, its 320 billion parameter, 18 billion active, 1 million token context model, is now deployed on more than 100,000 Chinese accelerators, with end-to-end throughput tripled in under two weeks and per-token cost and hardware efficiency it claims are comparable to mainstream Nvidia GPUs. Huawei rotating chairman Eric Xu said Chinese AI providers may need to speed up their pace to perceive frontier risks, and forecast that agents will account for more than 90 percent of global AI processing traffic by 2035. A Zartbot analysis of DeepSeek V4.1 Flash attributes its roughly 420 tokens per second to three-axis KV cache compression across channel, sequence, and layer, cutting runtime KV storage to a quarter and persistent storage to an eighth of the prior V4-Flash.
Cost parity with Nvidia on domestic silicon is the claim, and it is the claim that export controls were designed to make impossible. Z.ai raised $5 billion on Tuesday with 60 percent earmarked for GLM-6 and infrastructure, JD Cloud demonstrated 100,000-unit domestic deployment last month, and now the flagship open model runs at scale on chips Nvidia does not sell. If the per-token number holds under independent load, the HBM shortage is the only constraint left, and Beijing's five-year plan named advanced memory as a priority on Tuesday.
Contrarian take: Xu telling Chinese labs to speed up in the same week Amodei, Nadella, and von der Leyen ask everyone to slow down is the pacing debate's real problem stated by the other side. The DeepSeek KV compression writeup is the engineering reason Flash is cheap, and it is a memory trick, which is the same lesson as Positron and Meta's MTIA: whoever needs the least HBM per token wins the next two years. The Kimi K3 review covers the other Chinese flagship in the CISA advisory.
GPT-6 Astra Scores 49.2 Percent vs Opus 5 at 28.8 on Microsoft's ProgramDistill
Microsoft researchers released ProgramDistill, which mines 4,063 software engineering tasks from 26 reference applications and asks models to reconstruct them; GPT-6 Astra reaches 49.2 percent on full reconstruction and Claude Opus 5 28.8 percent. A HarnessTax study benchmarked 21 model-harness pairs across Claude Code, Codex CLI, and Pi and found harness choice barely moves success rate but substantially changes token cost. Nvidia's Agora ran 13 language-model agents through a Git workflow for 12 days, producing 1,703 contributions that improved an evaluator metric from 3.39 to 1.899 bits per byte, closing 62 percent of the gap to a GPT-2 124M baseline. Cambridge's XConf confidence estimator beat 10-sample self-consistency on 23 of 24 AUROC comparisons and lifted selective-prediction agent success by up to 8.7 points.
A 20-point gap between Astra and Opus 5 on the same reconstruction task is the widest split between the two flagships on any coding benchmark this month, and it is on Microsoft's benchmark, which is worth remembering. HarnessTax is the finding that cuts against the week's narrative: Cognition, Sakana, and Factory are all selling the harness as the product, and this study says the harness decides your bill, not your outcome. Both can be true if the outcome is roughly fixed by the model and the harness is where the margin lives.
Agora is the small-scale version of the automation index: 13 agents, 12 days, 1,703 commits, and a measurable research result with no human writing code. Scale that by a thousand and you get Anthropic's 30,000 agents. The AI coding tools hub tracks the harnesses HarnessTax measured.
Edge0 Streams a 35B Model From SSD at 20 Tokens a Second on a Mac Mini
AutoArk's Edge0 keeps a 35 billion parameter class mixture-of-experts model on SSD and streams active experts into memory, reaching 20.4 tokens per second on a 24 gigabyte Mac mini M4 Pro with 2.9 gibibytes of active memory, against 3.9 tokens per second and 18.2 gibibytes for the baseline, with an average quality gap of 3.9 points at the 35B tier and 2.8 at 8B. Zing-0.5 shipped a 5 billion parameter playable world model at 24 frames per second and 832 by 480, scoring 81.0 overall and 88.5 on consistency on WBench Navigation at about $0.009 per stream-minute. ActionPiece, a vision-language-action tokenizer on Qwen3-VL-4B, scored 94.8 percent on LIBERO, 68.8 on LIBERO-Plus, and 71.9 on SimplerEnv. Shanghai AI Lab's SP3O fixes PPO value flattening by applying value loss to three separated states per response.
Running a 35B MoE on a consumer Mac with under 3 gigabytes resident is the local inference story of the month, because it means the memory constraint that is repricing every data centre chip can be routed around at the edge by treating SSD as the expert store. A 3.9-point quality gap is the price, and for most local tasks it is worth five times the speed. Zing at under a cent per minute is the same trick applied to world models: playable video at 24 frames per second is now a consumer-hardware workload.
What to watch: Edge0 lands the same week DeepSeek's KV compression analysis explains Flash's throughput, and both are memory engineering rather than model improvements. The next generation of local models will be judged on active gigabytes per token, not parameters, and Apple's Siri on Gemini, which runs partly on-device, is the first mass-market product that number applies to.
Novo Nordisk Adopts Claude for Drug Discovery as Emulate Raises $700M in a Month
Novo Nordisk announced it is using Anthropic's Claude models across drug discovery and R&D workflows, with chief executive Mike Doustdar saying it will supercharge the R&D organisation. Emulate, a one-month-old startup founded by former DeepMind researchers, is in advanced talks to raise $700 million at a $3.7 billion valuation co-led by Index Ventures and Lightspeed, among the fastest-scaling frontier-lab spinouts on record. Isomorphic Labs said on Tuesday its models stay locked in-house while it works with Novartis, Eli Lilly, and Johnson & Johnson.
Novo Nordisk is the largest pharmaceutical company in Europe and it chose the general-purpose lab over the DeepMind spinout built for the purpose, which says the buying decision in pharma is now about agentic workflow rather than a specialised structure model. It also lands the same day Anthropic published that Claude leads a quarter of its own research, which is the sales pitch in one number: if it runs Anthropic's R&D, it can run yours.
Honest assessment: $3.7 billion for a company that is one month old and has shipped nothing is the valuation the market puts on a DeepMind alumni list in September 2026, and it follows Discovery Loop's $50 billion ask and Mechanize's $1.5 billion talent deal in the same fortnight. The frontier is being priced per researcher, and the automation index says each researcher now comes with a model doing a quarter of the work. The Claude AI complete guide covers what Novo Nordisk is deploying.
Crusoe Raises $3.9B at $30.9B as Amazon Signs an $8B Generac Generator Deal
Crusoe raised $3.9 billion at a valuation near $30.9 billion, tripling from $10 billion in October 2025, to mass-produce its modular Spark AI data centres from facilities in Colorado and Tulsa. Amazon signed an agreement with Generac worth up to $8 billion for backup generators, with $2.4 billion of initial deliveries expected in 2027 and 2028, and received a warrant for about 1.69 million shares at $200.93, a 2.6 percent stake; Generac shares rose more than 40 percent. Scotland's parliament backed a de facto moratorium on new hyperscale AI data centres pending updated planning guidance, with more than 20 projects proposed.
Factory-built data centres are Crusoe's answer to the constraint every operator has hit this month, which is that the bottleneck is now construction time rather than chips, and a valuation that triples in eleven months says investors believe it. The Generac deal is the same story from the power side: Amazon is buying $8 billion of backup generation because grid connections are years away, and taking equity in the supplier so nobody else can buy the capacity first.
Scotland's moratorium is the first outright pause on hyperscale builds by a national legislature, and it comes the same week the AI Energy Management Alliance launched to make data centres grid-flexible. The alliance is the industry's argument against exactly this kind of vote. Anthropic's 2.16 gigawatt Queensland site, covered in yesterday's edition, is what an operator builds when the grid says yes.
House Votes 417 to 3 to Make Data Centers Pay Full Grid Costs
The US House passed the Ratepayer Protection Act by 417 votes to 3, amending the 1978 Public Utility Regulatory Policies Act to require large data-centre customers to cover the full cost of the grid upgrades they trigger rather than spreading them across residential bills. Only Summer Lee, Delia Ramirez, and Rashida Tlaib voted no. The Financial Times separately reported that 20 soldiers using AI targeting now handle the workload of 2,000 in the 2003 Iraq invasion, with programmes pushing toward 1,000 tactical decisions an hour, one every 3.6 seconds.
A 417 to 3 vote is as close to unanimous as the House gets, and it is the political consequence of consumer electricity bills rising with AI demand. The bill does not slow construction, it reprices it, and for hyperscalers already committing hundreds of billions the grid-upgrade line is a rounding error. For smaller operators and the neoclouds financing on bank debt, it is a real cost that lands on the same balance sheets that just borrowed $22 billion and $29.6 billion.
Why this matters: this is the narrow data-centre bill Speaker Johnson promised instead of a broader AI framework, and it passed in a week. The FRONTIER Act with its mandatory audits has three labs supporting it and no floor vote scheduled. Congress is regulating the electricity, not the models, and the FT's 3.6-second targeting cadence is a reminder that the models are already deployed where no framework reaches.
Modi Doubles India's Chip Mission to $13.5B With $5B From Applied Materials
Prime Minister Narendra Modi unveiled phase two of the India Semiconductor Mission with $13.5 billion over 12 years, up from $8 billion in phase one. Applied Materials committed $5 billion over ten years and Lam Research pledged about 10,000 crore rupees toward the first Indian fab. Shanghai Biren placed about $1 billion of shares in its third raise since January, and China's five-year plan on Tuesday set a 9,800 exaflop computing target with advanced memory as a priority segment.
Two equipment makers committing before a fab exists is the sequencing that matters, because tools, not subsidies, are what India lacked in phase one. Five billion dollars from Applied Materials is a bet that the fab gets built, and it is the largest single foreign commitment to Indian semiconductor manufacturing to date. The comparison with China's plan is the scale gap: $13.5 billion over 12 years against 3.8 trillion yuan of infrastructure by 2030.
For Indian AI builders the practical timeline is unchanged, since a first fab on phase-two money produces mature-node chips at the end of the decade, not accelerators. What changes now is the supply chain conversation with Nvidia, AMD, and the Chinese vendors, because a country with a $13.5 billion mission and Applied Materials on the ground is a country that gets allocation.
Brookings and Fudan Propose Nuclear Red Lines for AI 6 Days Before Trump Meets Xi
Researchers at Brookings and Fudan University jointly proposed explicit US-China red lines barring AI from autonomously deciding the use of nuclear weapons, plus a dedicated military hotline for AI incidents, ahead of the September 24 meeting between President Trump and President Xi. It follows Elon Musk's proposal on Monday for a standardised pre-release test harness run by US labs and three or four leading Chinese companies, and Huawei's Eric Xu urging Chinese labs to speed up on Wednesday.
A joint American-Chinese proposal on nuclear command is the one form of AI coordination that has precedent, because the Cold War produced exactly these instruments, and it is the one Beijing's foreign ministry did not reject when it called Amodei's export-control essay fearmongering on Monday. The hotline is achievable in a single meeting. The red line on autonomous nuclear decisions is the kind of thing both leaders can sign without conceding anything on chips.
What to watch: whether the September 24 readout mentions AI at all. If it does, the hotline is the likely deliverable and the standards body the labs are building becomes the domestic half of a bilateral structure. If it does not, Musk's cross-lab harness stays a tweet and the FINRA model has no international counterpart. Six days.
Pew: 56 Percent of Democrats Now Worry About AI, Ahead of Republicans at 49
A Pew Research Center survey of 3,488 US adults found 56 percent of Democrats are concerned about AI against 49 percent of Republicans, a reversal of the pattern in earlier surveys, with Democratic worry about AI job losses jumping 17 points to 75 percent. It follows a week in which Trump attacked Amodei for proposing a slowdown, Bernie Sanders proposed a superintelligence ban, and four House Democrats asked the Speaker to cancel recess until AI safeguards pass.
Party polarisation on AI has flipped, and the driver is jobs rather than safety. A 17-point jump in job-loss worry among Democrats in a single survey cycle is the public reading the same numbers this newsletter has carried all month: 42 percent of developers say AI writes half their code, Shopify merged 3,536 agent pull requests in 30 days, and Anthropic says Claude leads a quarter of its own research. The safety debate is happening among executives. The jobs debate is happening among voters.
Honest take: the labs are aligning on pacing at exactly the moment the public is aligning on displacement, and those are different conversations. A standards body that tests for bioweapon and deception risk will not answer a 75 percent job-loss worry. The politician who connects the two wins the 2028 primary, and Sanders has already started.
Strix Agent Finds a 3-Year-Old Admin Token Exposing Baseten's GitHub
A Strix security agent discovered an unauthenticated Harbor container registry containing a GitHub token from March 2023 with admin access to Baseten's repositories and its Homebrew distribution channel; the token was rotated within hours. Jevlike shipped an open-source option scorer that replaces token-by-token decoding with a single probability per option, claiming a roughly 100x speedup over a small decoder, 98 percent on synthetic menus and 26 percent on real Wikispeedia data. Spain's AEPD, reported on Tuesday, remains the only regulator to have logged a breach executed entirely by an agent.
An agent finding a three-year-old admin token that no human audit caught is the defensive mirror of every offensive agent story this month, from PaperCut's 395 organisations to the Spanish breach, and Baseten is an AI inference company, which means its own customers' model weights sat one token away from a poisoned Homebrew tap. Rotation within hours is the right response. The token surviving since 2023 is the finding.
Builder guidance: run an agent against your own registries and artefact stores before someone else's does, because OpenAI's credential-search incident on Tuesday and this discovery describe the same behaviour from opposite sides of the ledger. Jevlike's 26 percent on real data against 98 percent on synthetic is the week's reminder that option scoring, like LLM rubrics in the ImpossibleRubrics study, collapses when the menu is not clean. Coverage of the OpenAI incidents sits in yesterday's edition, the compute build in the September 16 edition, and DeepSeek V4.1 Flash in the September 12 edition.
Frequently Asked Questions
What is the top AI news today, September 18 2026?
Anthropic's R&D Automation Index, published September 17, reports that Claude leads 26 percent of Anthropic's own AI research and development as of August 2026, up from under 1 percent in February, with more than 90 percent of work at the collaborates level or above and none fully autonomous. The same day OpenAI, Anthropic, and Google confirmed plans for a FINRA-style standards body, OpenAI shipped Astra for Law, and Z.ai said GLM-5.3-Flash runs on 100,000 Chinese accelerators.
Is Claude building the next version of Claude?
Partly. Anthropic says Claude leads, meaning completes most of a task end to end under human supervision, 26 percent of its AI R&D work as of August 2026, and participates at the collaborates level or above in more than 90 percent. It is not operating fully autonomously in any measured area. About 30,000 research and engineering agents run simultaneously under monitors that block roughly 1 in 47,000 actions.
What is the Anthropic R&D Automation Index?
A prototype measurement published by the Anthropic Institute on September 17, 2026 that catalogues about 15,000 R&D tasks into a 542-node hierarchy, has a Claude judge assign each an automation level on Epoch AI's AL0 to AL5 scale, and weights by person-time. It reported 26 percent at AL4 (leads), above 90 percent at AL3 (collaborates) or higher, and 6 percent of AI R&D compute on safety in a sampled July week. Anthropic plans regular publication with third-party verification.
What is the AI standards body OpenAI, Anthropic and Google are proposing?
An industry-funded self-regulatory body modelled on FINRA, the US financial industry regulator, to test powerful AI systems before public release. Demis Hassabis proposed it in July 2026, OpenAI's Chris Lehane said the three labs have worked on it for weeks, and Sam Altman endorsed it on September 15. Cohere's Aidan Gomez called it a cartel and Senator Sanders said binding international rules are needed instead.
What is Astra for Law?
Astra for Law, launched by OpenAI on September 17, 2026, is a configuration of GPT-6 Astra tuned for legal research and document drafting. It is OpenAI's first vertical edition of the GPT-6 flagship and competes with legal AI products such as Harvey that are built on OpenAI's models.
Does GLM-5.3 run on Chinese chips?
Yes. Z.ai reported on September 17, 2026 that GLM-5.3-Flash, a 320 billion parameter model with 18 billion active parameters and a 1 million token context, is deployed on more than 100,000 Chinese accelerators, with throughput tripled in under two weeks and per-token cost it claims is comparable to mainstream Nvidia GPUs.
What is the Ratepayer Protection Act?
A bill passed by the US House on September 17, 2026 by 417 votes to 3 that amends the 1978 Public Utility Regulatory Policies Act to require large data-centre customers to pay the full cost of the grid upgrades they cause, rather than spreading those costs across residential electricity bills. Three members voted against it.
How much is India spending on semiconductors?
Phase two of the India Semiconductor Mission, announced by Prime Minister Modi on September 17, 2026, commits $13.5 billion over 12 years, up from $8 billion in phase one. Applied Materials committed $5 billion over ten years and Lam Research pledged about 10,000 crore rupees toward India's first fab.
Recommended Blogs
● AI News Today September 17 2026: 14 Biggest Stories
● AI News Today September 16 2026: 14 Biggest Stories
● AI News Today September 12 2026: 14 Biggest Stories
● Claude AI 2026: Models, Features, Desktop and More
● Best AI Models July 2026: Ranked by Use Case and Price
● GPT-5.6 Review: Sol, Terra, Luna Benchmarks and Pricing
● Kimi K3 Review: Benchmarks, Pricing, and K2 Comparison
Resources & Community
Join our community of 70,000+ AI enthusiasts and learn to build powerful AI applications! Whether you're a beginner or an experienced developer, Build Fast with AI helps you understand and implement AI in your projects.
● Website: buildfastwithai.com
● LinkedIn: Build Fast with AI
Agentic AI Launchpad 2026
A structured 6-week cohort program that takes you from AI basics to building and deploying real-world agentic AI systems. Includes live sessions, expert mentorship, project reviews, and a builder community network.
Ready to go from learning to building? Join the next cohort: Agentic AI Launchpad 2026
Free AI Resources
Access free tools, workshops, and micro-learning to keep building:
● AI Workshops: Free resources, upcoming events, and past recordings
● Unrot: Learn AI in 5 minutes a day (free micro-learning app)
● Gen AI Experiments: free cookbooks and notebooks on GitHub
The Trump-Xi meeting on September 24, the Senate's turn on the Ratepayer Protection Act, and OpenAI's first six-day incident report all land within the week. Follow Build Fast with AI so each recap reaches you before your standup.
References
● Measuring the pace of AI development (Anthropic Institute)
● Claude leads 26 percent of R&D (Unite.AI)
● Labs in talks on AI safety for weeks (TechCrunch)
● AI standards body proposal (TechXplore)
● Astra for Law and Apple settlement ruling (LLM Stats)
● GLM-5.3-Flash on Chinese accelerators (AI Weekly)
● ProgramDistill, HarnessTax, Agora and XConf papers (AI Weekly)
● Edge0, Zing-0.5 and ActionPiece (AI Weekly)
● Novo Nordisk and Claude (Novo Nordisk)
● Crusoe $3.9B round and Amazon-Generac deal (AI Weekly)
● Ratepayer Protection Act vote (AI Weekly)
● India Semiconductor Mission phase two (AI Weekly)
● US-China AI nuclear red lines (Brookings)
● Pew survey on AI concern by party (Pew Research Center)
