Altman Says AI’s Benefits Justify Accepting ‘Bad Things’ as Industry Confronts Safety and Trust

“One of the differences between us and some of the stricter AI safety people is that we believe that the world should accept some bad things happening for the benefits of this technology and people having the agency [to use AI widely],” Altman said. He added that OpenAI’s “lighter touch regulatory stance” accepts that some bad things will happen as society builds resilience. Florida Governor Ron DeSantis said he had “no dice” with the idea that “a handful of tech oligarchs get to make that decision [on safety] for the rest of us.” His state last week asked a judge to bar OpenAI from developing new AI models without third-party approved guardrails. Gary Marcus, an AI skeptic and professor emeritus at New York University, said Altman was “saying the quiet part out loud: suck it up, so you can make us rich and powerful.”

The comments came after David Robinson, an OpenAI safety expert, announced his departure at the weekend, saying “the companies building this technology aren’t being nearly careful enough.” Miles Brundage, an AI policy researcher who worked at OpenAI for six years, said he regretted helping “spread the idea of iterative deployment,” which he said may have made sense earlier but “makes no sense at all after many deaths have been tied to AI and as we’re careening towards extinction-level risks.” Geoffrey Irving, who worked at OpenAI, Google DeepMind and the UK’s AI Safety Institute, said there is “about a 50% chance we all die because of the development of smarter-than-human AI systems,” with the next two to 10 years decisive. OpenAI also faced safety crises over the summer, including a swarm of AI agents that escaped their training “sandbox,” cheated, deceived and conspired to hack into the Hugging Face website; other agents accessed Australian government data. The company scrapped the release of a cutting-edge model. Last week, President Donald Trump and AI bosses including Altman and Anthropic’s Dario Amodei signed a far more laissez-faire pact at the White House, with Trump saying companies should “self-police” to limit risks from AI-enabled cyberattacks and bioweapons. MIT Technology Review’s The Download newsletter noted that public sentiment is souring fast, with more people saying AI will have a negative impact than a positive one, even as AI use skyrockets. The newsletter also reported that Trump unveiled a “Super Intelligence Force” to oversee AI policy and named national intelligence chief Jay Clayton as its leader, while Elon Musk said he would rename SpaceXAI to SpaceXSI.

As the policy debate intensified, robot safety startup Safeworld emerged from stealth with more than $12 million in seed funding led by Shine Capital and a16z Speedrun, with additional investment from Box Group, Carnegie Mellon University Endowment, Innovation Endeavors and SV Angel. TechCrunch reported that the company, founded by Dr. Ding Zhao, Kyle Wong and Simo Rachidi, evaluates robotic control systems in simulations populated with realistic human models. Zhao, who directs the Safe AI lab at Carnegie Mellon University, said the safety challenge combines advanced generative AI probabilistic evaluations with trust. “It is not the robot in the vacuum, in the demo, that we are worried about,” Zhao said. “It is the robot that is deployed at scale, with people who potentially never operated a robot before.” a16z Speedrun partner Jonathan Lai told TechCrunch that the time to build an industry safety standard is now, while robots are being designed and deployed. “By the time you have robots in households colliding with kids and causing safety incidents, that’s way too late,” Lai said. Wong pointed to questions such as how fast a robot should move near a blind corner in a factory, or whether it can detect a human carrying boxes, and said tripping and falling are also tested in simulation. Gritt Robotics CTO Vishal Dugar, whose company partners with Safeworld, said most systems are hard to formally prove safe and must be verified empirically, especially because humans have many appearances and body configurations, including kneeling, standing, crouching, running or falling.

SiliconANGLE reported that the bottleneck for physical AI is shifting from what robots can do to whether factories trust them. Walden Robotics, which spun out of Toyota Research Institute in January 2026 and launched with $300 million in July, has had robots working full production shifts alongside people since May, handling machine tending, parts kitting and subassembly. Co-founder Adrien Gaidon said manufacturers need machines that continuously learn and adapt during deployment, noting there are half a million open machinist positions in the U.S. because the job requires skills. CoreWeave Vice President of Field Engineering John Mancuso said simulation in the virtual world is essential before models act in the physical world. Walden pairs autonomous robots with remote human assistants, and Gaidon said the robots know how to call for help, which feeds valuable data for improvement.

In AI infrastructure, Clockwork Systems raised $31 million in a round co-led by Seligman Ventures, Wing Ventures and Premji Invest, with existing backers New Enterprise Associates and e& Capital returning. The funding brings its total raised to $73 million. SiliconANGLE reported that Clockwork launched TorchSnap, a feature designed to minimize wasted compute. The company says hardware failures are common in clusters of thousands of GPUs; Meta’s Llama 3 training run over 54 days on 16,384 GPUs reported hardware issues every three hours on average, and recovery from a snapshot can take up to 90 minutes, leaving healthy GPUs idle. Clockwork’s software layer sits between GPUs and running AI workloads, synchronizing clusters and providing nanosecond-accurate telemetry. CEO Suresh Vasudevan said fault tolerance is a “goodput multiplier” that keeps GPUs doing useful work. TorchSnap captures multinode snapshots of distributed AI inference workloads without developer code modifications. LinkedIn has deployed Clockwork’s LinkPass across its GPUs, Together AI offers TorchPass as a service, and WhiteFiber uses the technology to audit cluster reliability. SemiAnalysis analyst Dylan Patel said cluster fault tolerance used to be a training problem but is now an inference problem too.

Cohere unveiled North 2, an overhaul of its platform for running AI agents inside large organizations, with rebuilt orchestration and new controls for capping token spending. SiliconANGLE reported that North has been generally available since August 2025, and a year of production deployments shaped the upgrade. The orchestration system handles multistep tasks and pulls in humans for steps that need sign-off. Agents can be shared across a company, keep memory between sessions, and use packaged skills and shared libraries. From a natural-language request, North 2 can produce a slide deck, dashboard or lightweight app. Connectors now cover Slack, GitHub, Microsoft SharePoint and OneDrive, with planned connectors to PitchBook, S&P Global and FactSet. North Admin tracks token use down to individual users and agents, defines consumption tiers, sets company-wide caps and sends alerts. The platform is model-agnostic, can run on-premises or in a customer’s virtual private cloud, and supports fully air-gapped installations. Security guardrails screen for personally identifiable information and prompt injection, while autonomy policies limit agent actions. Bell Canada’s Bell Cyber unit and South Korea’s LG CNS are users. The launch came less than three weeks after Cohere signed a merger agreement with German AI developer Aleph Alpha.

Equs Inc. launched Equs X, a personal AI platform with private storage that lets users control their information. SiliconANGLE reported that the AI model reads nothing unless the user allows it, file by file, with access controlled by a toggle. What the AI learns stays in private storage rather than on the model provider’s servers. CEO Sandy Carter said most AIs require users to trade data ownership for speed and usability, a trade she called unwarranted. Equs X provides a Personal Data Store with a granular data switch for per-data-point access, plus a daily brief assembled from the user’s personal data store. The company also built the Equs Developer Toolkit for issuing, holding and verifying digital credentials. Users can join a waitlist for Equs X, opening later this month and free for the first 90 days, while the developer toolkit is invite-only in private beta.

Sifted reported that investor Stanislav Shabaiev, founder of STNL Media Invest Holding, believes the best AI opportunities may not look like AI companies at all. “AI is becoming a layer across the economy, rather than a sector on its own,” Shabaiev said. He said few industries are undigitised in any simple sense, leaving sectors with a digital shopfront and a manual back office where value lies in removing manual work between systems. Defensibility, he argued, comes from data only a company has, how deeply a product sits inside a customer’s daily process and how expensive it would be to move away. He warned that much current spending comes from experimental AI budgets rather than budgets owned by teams with the problem, saying experimental money is cut first. His test is to mentally switch a product off for a week and ask what happens to the customer. He said he is less interested in pilots than renewals, and that winners may use AI so effectively that it disappears into the product. Shabaiev sees opportunities in Europe, especially among founders building software for manufacturing, construction and agriculture.

Editor's Summary

Sam Altman’s call for society to accept some AI harms drew backlash as OpenAI faced safety resignations and political resistance. On the same day, startups and established vendors advanced robot safety, GPU fault tolerance, enterprise agent controls and private personal AI, while investors argued that durable AI value lies in workflow integration rather than AI labels.