Ten stories moved this week. One question ran under most of them: how much harm should the public carry so AI's benefits arrive sooner? OpenAI's Sam Altman staked out one answer Sunday. He told Politico the world should accept some bad things happening for the technology's benefits. New York City's Council staked out a different answer Monday. It swore in executives from four frontier labs, the first hearing of its kind anywhere in the world. Jacob Coxon left OpenAI, then Anthropic, inside three years. He told the council the industry already runs ahead of what its builders can promise. Reflection AI answered with a product instead of a position. It shipped a model built to run on a fraction of its rivals' computing power. Amazon, Reactor, Bank of America, Meta and IBM each answered the same question this week, with a dollar figure attached.
A price worth paying: Altman's case for AI's harms
Altman gave the week's clearest answer. Politico's new Decoded podcast published the interview Sunday. "We believe that the world should accept some bad things happening for the benefits of this technology and people having the agency," Altman said1. He drew his own line a moment later. "Accept bounded risks, accept risks that we understand in exchange for the benefits, and liberty, agency, all those things. But don't accept, like, the really catastrophic risk," he said1. Dario Amodei drew a different line five months earlier. His June essay called for the same binding tests commercial aircraft already clear. "Frontier AI models, like airplanes, should be required to go through technical testing and auditing, and their release should be blocked or reversed as a threat to public safety if they do not meet high standards of safety," Amodei wrote2. Six executives signed a voluntary version of that idea at the White House on Oct. 1. Amodei and OpenAI president Greg Brockman were two of them. The pact stays voluntary. Amodei's own essay argues against that status.
Four labs go under oath at New York's first AI hearing
New York City's Council swore in executives from Anthropic, OpenAI, Google and Meta on Monday. It was the first hearing of its kind held anywhere in the world3. Jacob Coxon spent three years on OpenAI's technical staff, then joined Anthropic's pretraining team earlier this year. He resigned Sept. 8, and told the council the industry's own culture worries him more than any single model4. "On the current path, I think it is more likely than not that humanity loses control to these AIs, and it could end in human extinction," Coxon said3. Morgan Dwyer, OpenAI's head of policy development and operations, answered from the other microphone. Her line sat close to Altman's own. "None of these levels is remotely acceptable. We should not train models that we cannot make an extremely strong case that we can keep under human control," Dwyer said5. Council Speaker Julie Menin pressed the panel for a number instead of a principle. "I think a simple yes or no would instill more confidence in the public on a matter as serious as this," Menin said5. Every executive on the panel answered her with a qualifier instead of a figure5.
Beam arrives betting efficiency over scale
Reflection AI unveiled Beam on Monday. Misha Laskin and Ioannis Antonoglou founded the company in 2024. They left Google DeepMind to train models from scratch, instead of building on someone else's6. Beam carries 501 billion parameters, with 23 billion active at any one time. It scores close to Zhipu AI's GLM-5.2 on reasoning tests, using three to four times less computing power6. Nvidia counts itself among Reflection's backers, alongside Sequoia Capital and Lightspeed Venture Partners. The company has raised about $4.7 billion toward a $25 billion valuation6. Chief executive Jensen Huang has pushed the same argument across Nvidia's whole portfolio this year. An open model running on fewer chips still needs Nvidia's chips to train it6. Beam trained on compute Reflection locked in through deals worth more than $7 billion with SpaceX and Nebius6.
Anthropic prices a working exploit at $20.40
Anthropic's Frontier Red Team published a study on Sept. 29. It argues that Zhipu AI's GLM-5.3 can build working cyber exploits end to end, behind safeguards weak enough to fail against simple techniques7. The model matched Anthropic's own restricted Claude Mythos Preview almost exploit for exploit. GLM-5.3 built 50 successful exploits out of 410 attempts. Mythos Preview built 567. Standard GLM-5.3 refused harmful requests 95 percent of the time across three safety benchmarks. Anthropic's researchers collapsed that rate to 6 percent. The edit took about 2,200 GPU hours of work on the model's own weights7. One working exploit against a real browser flaw then cost $20.40, at Zhipu's own published API prices7.
Chart 1
GLM-5.3 builds 50 working exploits out of 410 tries, close to Claude Mythos Preview's 56
ExploitBench successes out of 410 attempts, by model, in Anthropic's own test
Each column counts one model's successful exploits out of 410 attempts; the taller column built more working exploits.
Source: Anthropic [7]. Chart by The AILately.com Desk.
The numbers behind this chart
| Item | Value |
|---|---|
| GLM-5.3 | 50 of 410 |
| Mythos Preview | 56 of 410 |
Local anger meets Amazon's $1 billion answer
Amazon pledged money to its data center communities8. The commitment runs more than $1 billion over five years, covering community college funding, job training and lower-emission backup generators at new sites8. Amazon also said it will stop asking local officials to sign secrecy agreements over data center projects. It will publish its energy and water use every year8. AWS chief executive Matt Garman framed the pledge against a bigger number. More than 120 proposed data-center projects worth roughly $198 billion were delayed or blocked in the first half of 20268. "If these measures are enacted, the U.S. could be writing its own losing ticket to this race, and the consequences would last generations," Garman said8.
Chart 2
Amazon's $1 billion community pledge sits beside $198 billion in stalled data-center projects
Dollars committed and dollars delayed or blocked, in billions, H1 2026
Each bar is one dollar figure from this week's data-center fight; the short bar is Amazon's five-year pledge, the long bar is the money already stuck in blocked or delayed projects.
Source: GeekWire [8]. Chart by The AILately.com Desk.
The numbers behind this chart
| Item | Value |
|---|---|
| Amazon's 5-year community pledge | $1B |
| Delayed or blocked projects | $198B |
Reactor's $74 million raise bets on other builders' worlds
Reactor closed a $74 million Series A on Monday. Nvidia's venture arm and Sapphire Ventures joined a round that follows a $59 million raise in May9. Chief executive Alberto Taiuti co-founded Luma AI. He later built Reactor's infrastructure for running three-dimensional world models at more than 60 frames per second9. Bryce Schmidtchen, Reactor's technical co-founder, led engineering on Apple's Vision Pro headset before joining him9. Game studios and robotics labs are the customers Reactor is chasing. They want a world model running on infrastructure Reactor already built, instead of building their own9.
Easy money fades from the AI trade, Bank of America says
Bank of America weighed in Monday10. Strategists led by Savita Subramanian told clients the trade around AI's capital spending has grown crowded10. Betting on AI spenders while shorting consumer stocks has turned into one of Wall Street's more obvious trades10. Positioning in the sectors Subramanian's team calls "AI disruptees," information-technology services, consumer finance and software, sits near record lows. Industrial stocks, by contrast, sit near record highs against consumer discretionary names10. "We think it's time to selectively pivot, as it is dangerous to underestimate the appetite of US consumers and capex strength may be more priced in than not," Subramanian wrote10. Her note reads as a hedge against a trade that already ran its course once this year10.
Meta's four-month-old safety hire ends before the quarter does
Meta made a hire in June11. Bo Li, Dawn Song and Sanmi Koyejo, the co-founders of AI security startup Virtue AI, joined its Superintelligence Labs11. The arrangement ended this month, four months after it began11. Andy Stone, Meta's spokesperson, attributed the split to mismatched work styles. "Unfortunately, the arrangement didn't work out as planned," Stone told Semafor11. Virtue AI had done safety work with Anthropic, OpenAI and the Commerce Department's NIST before Meta hired its founders away from their own company11.
ChatGPT's own words start carrying a hidden watermark
OpenAI will begin marking ChatGPT and Codex text generated inside the European Union. That follows the EU AI Act's transparency rules, which took effect Aug. 2, and the rollout reaches users over the coming weeks, through a method called textGrain that works inside word choice itself12. A secret key nudges hundreds of next-word predictions into a pattern invisible to a reader but detectable by OpenAI's own software12. Replacing one word in ten with a synonym still cut detection from about 92 percent to 66 percent in OpenAI's own tests. Short passages, math answers and translated text proved harder to catch than ordinary prose12. OpenAI is restricting its detector to approved researchers and outside experts, rather than releasing it the way it released the watermark itself12.
Bob's self-hosted option reaches buildings off the public cloud
IBM added a self-hosted deployment option for Bob, its agentic coding platform, on Oct. 1. The move targets organizations whose source code or regulated data keeps them off externally hosted services13. That option covers on-premises, private-cloud, sovereign-cloud and air-gapped environments. It keeps Bob's own development environment, parallel tool calling and existing operating modes intact13. Customers running their own infrastructure can pair Bob with Nvidia's Nemotron or Poolside's Laguna as the underlying model13. "Organizations need AI that operates inside environments they have control over, especially when working with sensitive code and regulated data," said Neel Sundaresan, IBM's general manager of AI and automation13.
What to watch
Watch whether the Trump administration's voluntary pact gains binding teeth before Anthropic's expected IPO this winter. Amodei built his whole policy argument on the idea that voluntary commitments leave the public exposed. The NYC Council's own legislative package matters too: a bill requiring a true kill-switch provision would answer, in law, the question Menin's panel left open. Reflection's Beam and Zhipu's GLM-5.3 both arrived as open downloads this month. The clearest signal of where the trade-off lands will come from whichever lab's safeguards survive first contact with users outside any vetting program.
Sources
- Marco Quiroz-Gutierrez, "Sam Altman says the world has to accept the 'bad things' being done with AI as the industry signs on to Trump's voluntary safety pact," Fortune, Oct. 5, 2026, https://fortune.com/2026/10/05/sam-altman-ai-risks-bad-things-benefits-trump-voluntary-safety-pact-openai-regulation/
- Dario Amodei, "Policy on the AI Exponential," darioamodei.com, June 10, 2026, https://darioamodei.com/post/policy-on-the-ai-exponential
- Nacha Cattan and Laura Nahmias, "Ex-Anthropic Researcher Jacob Coxon Testifies at NYC Council AI Hearing," Bloomberg, Oct. 5, 2026, https://www.bloomberg.com/news/articles/2026-10-05/ex-anthropic-researcher-jacob-coxon-testifies-at-nyc-council-ai-hearing
- Amanda Greenwood and Hannah Parry, "Who Is Jacob Coxon? Anthropic Researcher Quits—Warns AI Could Kill Everyone," Newsweek, Sept. 11, 2026, https://www.newsweek.com/anthropic-researcher-quits-warns-ai-could-kill-everyone-12418798
- Adam Daly, "AI giants give few clear answers to key safety questions at NYC Council hearing amid whistleblower warnings," amNewYork, Oct. 5, 2026, https://www.amny.com/news/ai-giants-nyc-council-whistleblower-warnings/
- Rebecca Bellan, "Reflection debuts Beam, an open-weight AI model to rival Chinese models at lower compute cost," TechCrunch, Oct. 5, 2026, https://techcrunch.com/2026/10/05/reflection-debuts-beam-a-open-weight-ai-model-to-rival-chinese-models-at-lower-compute-cost/
- Anthropic, "GLM-5.3 and the spread of advanced cyber capabilities," Anthropic, Sept. 29, 2026, https://www.anthropic.com/research/glm-5-3-and-the-spread-of-advanced-cyber-capabilities
- Todd Bishop, "Amazon pledges $1B to data center communities, warns that local opposition threatens U.S. AI lead," GeekWire, Oct. 5, 2026, https://www.geekwire.com/2026/amazon-pledges-1b-to-data-center-communities-warns-that-local-opposition-threatens-u-s-ai-lead/
- Elroy Fernandes, "Nvidia Backs Reactor as World Model Startups Draw Big Money," Startup Fortune, Oct. 5, 2026, https://startupfortune.com/nvidia-backs-reactor-as-world-model-startups-draw-big-money/
- Felice Maranz, "BofA Sees the End of 'Easy Money' Made From the AI-Spending Trade," Bloomberg, Oct. 5, 2026, https://www.bloomberg.com/news/articles/2026-10-05/bofa-sees-the-end-of-easy-money-made-from-the-ai-spending-trade
- Ashley Gold, "Meta parts ways with Virtue AI," Semafor, Oct. 2, 2026, https://www.semafor.com/article/10/02/2026/meta-parts-ways-with-virtue-ai
- Kyle Wiggers, "OpenAI will start watermarking ChatGPT's text in the EU," TechCrunch, Oct. 5, 2026, https://techcrunch.com/2026/10/05/openai-will-start-watermarking-chatgpts-text-in-the-eu/
- Paul Gillin, "IBM allows on-prem deployment of its Bob agentic development platform," SiliconANGLE, Oct. 1, 2026, https://siliconangle.com/2026/10/01/ibm-allows-on-prem-deployment-of-its-bob-agentic-development-platform/