Sam Altman: ‘I No Longer Think We’re Gonna Be Able to Just Patch All the Bugs’
THE SO WHAT
When a leading lab CEO says we won’t be able to “just patch all the bugs,” he’s telegraphing that AI risk management will look more like safety engineering than app updates. Operators should be budgeting for continuous red-teaming, layered controls, and incident response around models—not assuming vendor patches will close every gap.
READ THE SOURCE
MORE FROM THE WIRE
Applied AIUS Government issues order to remove all DeepSeek products and third-party services developed using DeepSeek immediately, as more fears of Chinese LLM loom large
A blanket federal ban on DeepSeek is a clear line: provenance and jurisdiction of LLMs are now procurement criteria, not afterthoughts. If you sell into regulated or public-sector markets, expect hard questions about where your models were trained, who can access telemetry, and how you’ll swap vendors under pressure.
Applied AIAbliteration Launches Cheaper Version of Its ‘Unrestricted AI‘ at the Worst Possible Time
Cheaper “unrestricted AI” products widen the gap between what’s technically possible and what’s policy-compliant. If your platform can be a conduit for these tools, assume regulators and insurers will expect you to prove you’re not the soft target in the chain.
Applied AIOpenAI to Train AI Models Without Internet After Cyber Incidents
Cutting internet access during training after prior breaches is a clear signal that frontier model development now assumes adversarial conditions by default. If you’re training or fine-tuning sensitive models, air-gapping and strict egress controls just moved from “paranoid” to baseline.
Applied AIJack Dorsey’s Block made an AI model that talks to you like an awkward new friend
Block training Dessa “from scratch on nothing but human conversation” is a bet that tone and social nuance—not raw IQ—will differentiate assistants. If you’re building consumer-facing agents, your moat may be in how they make people feel, not just how fast they answer.