Top items
- Political AI ads this US midterm cycle top $137 million, with ~70% carrying no AI disclosure; a pro-AI super PAC paid $20M to a Nevada “company” that is only a PO box.
- 81 of 134 AI models failed a terrorism-prompt safety test; stripped open-weight models now appear online in under 3 days on average.
- DeepSeek V4.1-Flash-1017 released: 552B-param MoE with live web search, 1M context, MIT license, roughly on par with Opus 5.
- Claude (Anthropic) autonomously redesigned 8 quantum circuits (up to 17× fewer gates) and solved a 20-year NIST elliptic-curve seed mystery in 30 minutes.
- Anthropic’s own data shows AI-agent work hours surpassed human hours (6:1 ratio) and junior hiring collapsed.
- First large study of doctors using AI counts 270 million questions sent to OpenEvidence.
Policy & elections
AI-generated midterm ads exceed $137M, mostly unlabeled. Campaigns and PACs have spent more than $137 million on AI-generated or AI-enhanced ads this US midterm cycle, according to Wesleyan Media Project and AdImpact data reported by the New York Times. About 70% of the ads reviewed did not disclose to viewers that AI was used. The biggest pro-AI super PAC network, Think Big (part of the Leading the Future network), paid about $20 million to Lantern Production Consultants — a company Wired found exists only as a post office box in Sparks, Nevada. The network has raised over $140 million, with backers including Marc Andreessen, Ben Horowitz, and OpenAI president Greg Brockman. The Campaign Legal Center filed an FEC complaint in May, but the FEC has lacked the quorum needed to rule since May 2025. Separately, five national polls in 2026 found that 55%–77% of US adults say AI is moving too fast or should slow down; in three of the four polls that broke results out by party, a majority of Republicans also said AI was moving too fast. This week’s AI Weekly poll asks whether AI-made political ads should be legally required to disclose AI use.
Safety & security
81 of 134 AI models failed a terrorism-prompt safety test. Tech Against Terrorism fed 134 models more than 200 prompts on topics like mass-casualty attacks, and only two passed completely, Wired reported. The group also found that copies of new open-weight models with their safety limits removed now appear online in under three days on average, down from more than 24 days in 2024. As an illustration, Meta’s Llama 3.1 8B scored 97 out of 100 on the test, but a version with its limits removed scored just 3. The takeaway stressed: assume any open-weight model can be stripped of its safety guardrails within days.
Police posing as AI-generated people to catch suspects. Police in Livonia, Michigan used an AI-generated photo of a teenage boy to catch a man suspected of seeking sex with minors, per an affidavit reported by 404 Media on October 16. An FBI task force officer created a profile named “Caleb” on the dating app Taimi using the AI image and listing no age. The suspect messaged “Caleb,” the chat moved to WhatsApp, and when the officer said Caleb was 15, the man allegedly sent sexually explicit messages. Investigators then used Clearview AI, Thomson Reuters CLEAR, and Flock to identify him and track his car to the meeting place, where he was arrested. This builds on prior reporting: in April 2025, 404 Media reported that New York company Massive Blue sells “Overwatch,” a product that runs lifelike AI personas over text messages and social media, targeting “college protesters,” “radicalized” political activists, and suspected traffickers, with departments near the US-Mexico border paying hundreds of thousands of dollars. The Livonia affidavit shows one department’s use in one case but reveals nothing about how often police use AI personas or what rules govern them. Livonia police did not respond to 404 Media.
Field & industry developments
Companies that rushed into AI are pulling back. Meta scaled back plans to cut staff and lean on AI agents after employees spent growing amounts of time cleaning up technical and security errors, the Guardian reported (citing Reuters). Firms from Microsoft to McKinsey are now limiting how much employees use AI, and an Atlanta Federal Reserve working paper found 90% of executives say AI has not yet raised productivity at their companies.
First large study of doctors using AI counts 270 million questions. About 950,000 US physicians, physician assistants, and nurse practitioners sent 270 million questions to the AI tool OpenEvidence over roughly two and a half years, according to a Harvard Medical School study published in JAMA. Of the questions, 51.9% were about clinical decisions (mostly treatment) and 13.9% about diagnosis. A caveat: OpenEvidence supplied the data, and three co-authors are or were its employees.
Video game decompilation community shutting down because AI made it too easy. decomp.dev, a site where people decompile video games, is closing. Its admin said AI now accelerates decompiling so much that “the community aspect of decompilation has been broken.” A full decompilation of a mid-sized game used to take years.
Swedish TV production loses an entire season, sources blame an AI agent. The upcoming season of The Challenge Sverige was deleted after production company Mastiff gave an AI agent access to its server, Futurism reported, citing sources quoted by Aftonbladet. TV4’s program director disputed that AI was to blame, said the backup had failed, and said the raw footage still exists.
Anthropic’s own employment data: agent hours outnumber employees 6:1. Anthropic published internal data showing the share of junior-level staff in technical departments slipped from 21% to 14%, and junior hiring collapsed from 27% to 9%. Engineers increased code commits roughly 5× in six months, and interactive AI-agent work hours surpassed total human hours for the first time in March, reaching a 6:1 agent-to-employee ratio by August. Hiring criteria have shifted toward “autonomy and judgment,” with AI-tool-proficiency mentions tripling for technical roles and quadrupling for non-technical ones.
YC startup wants an AI agent watching every customer. Outlit’s founders say “every company will soon assign an AI agent to every customer to watch their every move,” pulling from call logs, sales data, social posts, and LinkedIn. The product has not been publicly released.
Tooling & model releases
DeepSeek V4.1-Flash-1017. DeepSeek posted V4.1-Flash-1017 to Hugging Face under an MIT license on Oct 17, 2026. It’s a sparse-MoE agent model with 552B total parameters that activates 8B on input and 16B on output, and supports a 1M-token context. The release adds Search and Fetch APIs for live web retrieval, priced at $0.75 per 1,000 successful requests with 5 RPS per account, and shrinks the KV cache to roughly 25% of the prior Flash. Reported benchmarks: 77.0 on DeepSWE v1.1, 42.6 on HLE, and 92.5 on BabyVision — placing it roughly on par with Opus 5 while beating GLM-5.3 and GPT-5.6 Sol.
Trending apps and models (“In the Wild”). Meta’s new open model is appearing in US search trends — Google’s trending “artificial intelligence news” searches pointed to it this week. Meta’s 800-billion-parameter Muse Spark Open model needs about 1TB of RAM to run locally, meaning two 512GB M5 Ultra Mac Studios or four 256GB ones, at roughly $38,000 for the four. xAI’s Grok Bot app is climbing the Productivity chart; it lets you message AI agents from your phone like coworkers, and the agents sign in to your tools/websites and keep working after you close your laptop. Summary, an app that records, transcribes, and summarizes meetings, interviews, lectures, and podcasts, is also rising. LocalForge, new on the Graphics & Design chart, runs Stable Diffusion image generation on an iPhone with no internet connection and no per-image fees (requires ≥6GB RAM). CHAI, climbing in Lifestyle, lets you chat with a large selection of AI characters that reply with their own voices, and create your own.
Research papers
Claude redesigns 8 quantum circuits, cutting gates up to 17×. Anthropic published results from an experiment where Claude worked largely autonomously to redesign quantum circuits for eight computational tasks, beating previously published algorithms. One modular-inversion variant uses 12% more qubits but 17× fewer gates. Claude also redesigned the Reed-Solomon decoder central to Decoded Quantum Interferometry, cutting gate count ~12× versus Google’s prior construction. On an iron-sulfide simulation, the model exploited molecular symmetries to compute results once and reuse them.
Claude cracks 20-year NIST ECDSA curve seed mystery in 30 minutes. Anthropic researchers say Claude discovered in under 30 minutes that NIST’s long-mysterious elliptic-curve seeds came from the C library’s lrand48 RNG initialized with Unix timestamps, rather than cryptographic hashing. P-256 took 1,842 iterations to produce its standardized seed; base points for six curves traced to the FreeLIP RNG with seeds between 20,000 and 20,517. The team says the finding increases confidence in the curves’ security because “anyone can now reproduce” the generation, and is donating a $36,864 bounty to the Internet Security Research Group. B-163’s seed and the base points for P-224, P-256, P-384, and P-521 remain unexplained.
Baidu Virtual Head Attention: 4× smaller KV cache. Baidu’s Virtual Head Attention paper, accepted to NeurIPS 2026, reports a 4× smaller KV cache than GQA-8 at dense scale. On a 30B MoE model it scored 2.60 points above GQA-4 across 12 benchmarks, and runs on a standard GQA-2 inference path with no new kernels. The largest test was 30B parameters.
RuleTrek: frontier models fail rule-following when rules sit in separate files. RuleTrek tested 11 frontier models on rules for tax, PostgreSQL, cybersecurity, software governance, procurement, and games. End-to-end scores ranged from 27% to 55% (at most 55% completion). A model that finds the correct rule often still fails to apply it.
PhD students offloading scientific writing to AI. An NBER paper measured AI writing in nearly every US STEM PhD dissertation from 2019 to 2026. It found none before 2023 and AI writing in 29% of dissertations filed in 2026. Graduates whose dissertations contain AI writing are less likely to go into academic research.
ML models poorly predict life outcomes. Matthew Salganik’s team, using a long-running survey, ran thousands of models per outcome and found that most life outcomes were not very predictable. Even for the most predictable ones, complex models showed little skill over a simple baseline method.
Miscellaneous (“Wait, What?”)
Many people claim to know AI executives who don’t exist. In a Harvard survey of more than 5,400 Americans, 53% of men said they knew of Mira Murati, Susan Milner, and Craig Dalton. Murati is real; the researchers invented Milner and Dalton. Among women, 40% said they knew the two fake names.
Palantir CTO’s smoking claim. Shyam Sankar said on CBS’s “Face the Nation” that “so many AI researchers have started smoking” to signal their faith that AI will cure cancer. He named no researchers and offered no evidence.