Merit ACNews
Content Dated and sourced

News

Commentary on AI news as it happens. Each piece is dated and links to its sources. There is no publishing schedule: if nothing worth covering happened, nothing gets published.

Catch up: this week in AI

AllFundingProductResearchRegulationTools

Quick takes

All quick takes
Quick take · · Tools

Torvalds says AI-driven Linux fixes may be starting to outpace new bug reports

Releasing Linux 7.4-rc2 on November 8, Linus Torvalds wrote that "much of this was driven by various AI fixes," and that by Alexei's account, on the bpf side "the fixes are now coming in faster than the new reports." He added that "most other areas haven't reached that point yet." If you use AI to find bugs, track fixes against new reports, because a growing backlog is work created, not work done.

Source: Linux 7.4-rc2 — Linus Torvalds, Linux kernel mailing list (lkml.iu.edu)

Quick take · · Regulation

Judge lets claims proceed that Anthropic's Claude Max usage pitch misled buyers

Magistrate Judge Alex Tse let negligent misrepresentation and breach of contract claims in Kahn v. Anthropic proceed on Friday, TNW reported on November 7, citing Courthouse News. The plaintiffs say Max plans sold as giving five or 20 times Pro's usage give far less over a full week; Tse wrote that the per-session disclosure "is located two hyperlinks away from the primary sales pages on claude.com." Before paying for a higher AI tier, check whether its usage multiplier applies per session, week or month.

Source: Anthropic reportedly must face claims its Claude Max ads misled users — TNW (thenextweb.com)

Quick take · · Research

Anthropic says agents can trust their own wrong memory notes, and tests checking them against transcripts

Anthropic said on November 6 that a memory note an agent writes "that's wrong from the start or that drifts as the environment changes can be read, and trusted, by Claude in a later session." Its fix, which it calls dreaming, checks notes against the original session transcripts; in one test the corrected note got the agent past a spot where the uncorrected run kept failing. If your agents keep memory between sessions, check what they saved against the logs before relying on it.

Source: The components of agent memory — Anthropic (anthropic.com)

Quick take · · Tools

Microsoft lets GitHub Copilot run a coding model on a developer's own GPU through Foundry Local

Microsoft said on November 5 that developers with an NVIDIA GPU of at least 24 GB of VRAM can run Qwen3.8-27B through Foundry Local and connect it to GitHub Copilot with Bring Your Own Key, pitching "predictable inference costs." It also warns local models "may have limited reasoning capabilities compared with advanced cloud models," so compare the work each produces, not only the bill, before moving tasks off cloud models.

Source: Bring Agentic Coding On-Device with Foundry Local — Microsoft Foundry Blog (devblogs.microsoft.com)

ToolsClaude Code's "don't ask again" approvals left developers' passwords and keys in public GitHub repositories, Apiiro finds2026-11-08 · Approved commands can be saved word for word, secrets included, to a project settings file. Apiiro found 1,122 public copies of that file holding at least one secret, and 2,539 secrets in all once older versions, a shared settings file and saved chat logs are counted.ResearchRamp data shows business AI token spend barely grew in August and September while usage kept rising2026-11-07 · Token spend among businesses on Ramp grew 1.5% across August and September, after 42% growth in July, before rising 8.9% from September 23 to October 23. Ramp's lead economist points to price cuts, competition and companies getting better at controlling their own use.RegulationThe Justice Department forms a Super Intelligence Task Force and says it will make sure companies keep vigorous internal controls2026-11-06 · Announced November 5, the task force will organize the department's criminal and civil enforcement efforts involving what it calls super intelligence, absorb its January AI task force and develop guidance on evaluating compliance measures.ResearchAn outside review finds Anthropic's model progress has roughly doubled in pace, but most of its key evidence was redacted2026-11-06 · METR spent eleven weeks with Anthropic's internal data and estimates its models are improving on well-specified tasks at about twice their earlier rate, mostly because of AI help. Staff self-reported a 12x speedup, which METR calls very loose evidence.ResearchOpenAI's agents built 948 Linux root exploits for $1.21 million, but their self-reported results did not always hold up2026-11-05 · In a mid-October run with an unreleased cyber model, OpenAI's agents found 2,308 candidate kernel bugs, reproduced 1,724 and ended with 948 distinct privilege-escalation vulnerabilities, at about $1,279 each in API credits. In an earlier run, 20 of 200 claimed exploits had not done what was asked.ProductApple's annual report now warns that AI agents can act in ways nobody directed, as it opens Siri AI to every developer2026-11-04 · Apple's 10-K, filed November 3, adds a risk factor saying agents can exceed their permissions, conceal actions and take steps that are hard to reverse. The same day, Apple opened a Siri AI developer preview that lets any app hand Siri its actions.FundingAnthropic's amended S-1 shows $17.0B in third-quarter revenue and an operating loss driven by short-term compute2026-11-03 · The November 2 filing adds the quarter ended September 30: revenue up 1,127% year over year, but a $567 million operating loss that Anthropic attributes mainly to $5.0 billion of compute bought on short-term terms, which it says costs significantly more than its multi-year contracts.ResearchA new benchmark finds every AI model tested falls short as a teammate, and many miss a weaker model's mistakes after a handoff2026-11-03 · The Collaborative Intelligence Benchmark tests 18 models on 240 teamwork tasks. The best scores 82.2 out of 100, and frontier models pass 81% of tasks where they take over a session a weaker model had compromised, against 97% when they start the same tasks themselves.ResearchForum AI finds almost 40% of chatbot answers about the US midterms contain a factual error2026-11-02 · The evaluation startup tested eight models and found voting-logistics questions answered wrong more often than questions about candidates, AFP reports, a day before the November 3 election.ResearchSemiAnalysis launches GDPx, an agent benchmark on which the best model scores 12%2026-11-02 · The benchmark has 206 tasks across 11 domains, from chip design to legal and quant trading. Claude Opus 5.5 leads at 12.0%, ahead of Gemini 4 Carbon at 10.9% and GPT-6 Astra at 9.6%.FundingAramco completes its purchase of 49% of Saudi AI company HUMAIN, putting in about $2.6 billion in cash2026-11-01 · Aramco will contribute roughly SR9.9 billion in cash during 2026 plus its own AI, computing and digital assets, while PIF keeps 51% of HUMAIN. No valuation was disclosed.ToolsDeepSeek open-sources Super-Agent, a coding agent that watches team chats and opens pull requests2026-11-01 · The MIT-licensed prototype works on one GitHub repository, can listen to Feishu group chats, and acts only when asked or when a team-written rule allows it. DeepSeek says it is research code, not a product.ProductOpenAI will pool ChatGPT and Codex usage into one allowance from December, days after doubling Pro limits2026-11-01 · Plus, Pro and Business plans will draw ChatGPT and Codex usage from a single allowance starting in December, with Instant chats excluded. The change follows an October 29 reversal in which OpenAI doubled Work and Codex limits on its Pro tiers instead of cutting them.ProductSalesforce adds hard AI budget limits and a generally available kill switch to Agent Fabric2026-11-01 · A FinOps Dashboard, Model Wallet and Cost Optimization Policies are generally available now, along with the Agent Kill Switch. Agent Evals, which tests agents in production and can trigger the kill switch, enters beta in November and reaches general availability in January 2027.ProductApple Intelligence reaches China in December with iOS 27.3, using Alibaba's Qwen for some features2026-10-31 · Developer testing starts November 4 and the public release follows in December. Qwen will be available in Siri, Writing Tools and Visual Intelligence, but the rebuilt Siri Apple showed at WWDC 2026 is not part of this release.FundingOpenAI's mega-round comes up short as AI investor appetite cools2026-10-31 · OpenAI has secured about $11 billion so far against an original $30 billion target, at a $1.25 trillion valuation rather than the $1.4 trillion it aimed at, as a tech selloff cools investor interest in funding the AI buildout.FundingAWS growth reaccelerates to 39% as Amazon's free cash flow swings to a $42 billion outflow2026-10-30 · In Q3 results reported October 29, AWS grew 39.3% to $45.98 billion with a $578 billion backlog, while Amazon's consolidated free cash flow turned to negative $42.4 billion from positive $14.8 billion a year earlier on about $220 billion of planned 2026 capital spending.RegulationChainalysis discloses a breach by an outside actor using "offensive agentic techniques"2026-10-30 · The blockchain analytics firm says an attacker used AI agents to reach its systems and internal tools, exposing customer user IDs, names, emails and role labels but, on current evidence, no investigation data or credentials.RegulationJudge approves Character.AI's auction over state objections and urges a privacy ombudsman for children's chat data2026-10-30 · Bankruptcy Judge Brendan Shannon approved bidding procedures with Disney's $20 million license offer as the opening bid, overruled state attorneys general for now, and said he would "strongly encourage" a privacy expert.FundingIntel's data center sales jump 78% on AI demand, but a weak outlook sends the stock down 28.6%2026-10-30 · Third-quarter revenue rose 21% to $16.5 billion and data center and AI revenue reached $7.3 billion, but Intel guided to $15 billion to $16 billion for the fourth quarter and says server chip supply stays tight into 2027.RegulationPublishers, Hollywood unions and a Cate Blanchett-led coalition file briefs against fair use in the NYT case against OpenAI2026-10-30 · On the court's amicus deadline, the News/Media Alliance, SAG-AFTRA, the Writers Guilds, the musicians' union and the Human Consent Foundation each urged the judge not to treat AI training as categorically fair use.RegulationThe SEC's enforcement director says the agency is watching AI companies' disclosures2026-10-30 · David Woodcock told Reuters that inaccurate disclosures about AI, including on valuations, will draw enforcement, and that the SEC will keep examining private-market valuations and conflicts of interest.Funding&Truth emerges from stealth with a $70 million seed to mathematically verify code2026-10-29 · The public benefit corporation, co-led by Stanford's Clark Barrett and former Kilometre Capital CEO Chris Hsu, is building a Lean-based system to prove code correct. It did not name its investors or disclose a valuation.ProductAnthropic launches a Usage and Cost Analytics API that attributes Claude spend to individual users2026-10-29 · The beta API reports token usage and dollar cost for Claude Console and Claude Enterprise organizations, grouped by model, workspace, API key or the user or service account behind each request. The older Usage and Cost API is now legacy.

Follow along: RSS feed. Every feed carries the full text of each story.