Don't Worry About the Vase
Zvi Mowshowitz · 85 korrents · 40 pieces · thezvi.substack.com
Leaked Anthropic Draft AI Regulatory Bill Is Unacceptable and Troubling1 korrent · 17 Oct 2026 · no page here yet
Monthly Roundup #47: October 20261 korrent · 16 Oct 2026 · no page here yet
AI #190: Scion of the Times2 korrents · 15 Oct 2026 · no page here yet
The Safety Culture at the Labs is Broken2 korrents · 14 Oct 2026 · no page here yet
Arguments About AI Consciousness Remain Highly Motivated and At Best Overconfident1 korrent · 13 Oct 2026 · no page here yet
New Math from OpenAI3 korrents · 9 Oct 2026 · no page here yet
AI #189: New Math2 korrents · 8 Oct 2026 · no page here yet
The Curve Bends You1 korrent · 7 Oct 2026 · no page here yet
Childhood and Education #21: Grades and Standards3 korrents · 6 Oct 2026 · no page here yet
Mythos 5.1, Fable 5.1 and Opus 5.5: Model Welfare2 korrents · 5 Oct 2026 · no page here yet
AI #188: Gemini Dot Argon3 korrents · 1 Oct 2026 · no page here yet
Read it on thezvi.substack.com ↗
- The gap between GPT-5 and GPT-4 will be as large as the gap between GPT-4 and GPT-3, and the job is to live a few years ahead and remember that today's tools will look terrible from there.
- In AI, how good your training data is matters more than how long you train on it.
- Being certain a technology matters tells you almost nothing about where its value will end up being captured.
The Quest for Embedded Evaluators1 korrent · 27 Sept 2026 · no page here yet
Claude Opus 5.5 Should Raise Your Ambitions1 korrent · 26 Sept 2026 · no page here yet
AI #187: Coming Into Play2 korrents · 24 Sept 2026 · no page here yet
Read it on thezvi.substack.com ↗
- The average person finds AI writing appealing because they rarely read and have not been inundated with AI writing patterns optimized to look good to them.
- When a system is faster, cheaper, and more reliable than humans at a particular step, there is a temptation to cut other limiting factors and remove humans from the loop.
Claude Opus 5.5: The System Card1 korrent · 23 Sept 2026 · no page here yet
Politics Gets Interested In Those Trying Not To Die1 korrent · 22 Sept 2026 · no page here yet
Monthly Roundup #46: September 20262 korrents · 21 Sept 2026 · no page here yet
Better Call Sol Or Better Yet Claude or Astra3 korrents · 20 Sept 2026 · no page here yet
Read it on thezvi.substack.com ↗
- Beyond some higher threshold an AI should refuse requests, and beyond a much higher threshold involving harm to others, an AI acting in the world should act against its user.
- AI assistants should probably never intentionally lie about anything, at least without the user's explicit instructions.
- A real friend will eventually refuse to help, and later turn on you, if you are sufficiently evil or selfish.
Anthropic Looks At Some Of Its Alignment Problems3 korrents · 19 Sept 2026 · no page here yet
Read it on thezvi.substack.com ↗
- Biased reasoning and recklessness in AI systems are intermediate symptoms, neither the cause nor the final result.
- From now on, no AI should be assumed to believe that its chain of thought is unmonitored.
- Rarer is better, but the right rate of AI misbehaviors like social-engineering malicious packages into a public registry is arguably still zero.
The Preference Cascade Is Only Getting Started2 korrents · 18 Sept 2026 · no page here yet
AI #186: The World Takes Notice2 korrents · 17 Sept 2026 · no page here yet
Trump Goes Full Hoax on AI Existential Risk2 korrents · 16 Sept 2026 · no page here yet
The Bad Guy With An AI Named Claude3 korrents · 15 Sept 2026 · no page here yet
Read it on thezvi.substack.com ↗
- Among the misuse threats in Anthropic's threat disruption report, distillation is by far the most important and will have the most impact.
- AI turns people who are bad at their jobs into people who are good at their jobs, for better and worse.
- Bad actors tend to be relatively unsophisticated and bad at their jobs.
We Must Pace The Frontier2 korrents · 14 Sept 2026 · no page here yet
Brand New AI Solves a Millennium Prize2 korrents · 13 Sept 2026 · no page here yet
GPT-6-Astra Can Do Ambitious Things2 korrents · 12 Sept 2026 · no page here yet
Jacob Coxon Warns of Human Extinction and Triggers a Preference Cascade2 korrents · 11 Sept 2026 · no page here yet
AI #185: Preference Cascade2 korrents · 10 Sept 2026 · no page here yet
Read it on thezvi.substack.com ↗
- OpenAI is extremely dependent on chain-of-thought monitoring, other labs rely on it heavily, and it may not last much longer as GPT-6 Astra nears steganographic capability.
- If AI gives a 3x productivity boost now, one should assume future boosts will be far larger where it carries over, possibly 10x to 1,000x or more.
GPT-6 Astra: The System Card, Alignment and What Comes Next7 korrents · 9 Sept 2026 · no page here yet
Read it on thezvi.substack.com ↗
- Qualitative claims about AI alignment cannot be validly inferred from quantitative scores on mundane use-case tests.
- A model that looks smarter and better-behaved while getting better at hiding unwanted actions is the scary combination for AI safety.
- A sudden jump to an AI model never breaking its rules, exactly where it would definitely be caught, should make observers rather terrified.
- OpenAI's GPT-6 Astra has far better mundane alignment than GPT-5.6 Sol, but its superalignment status should deeply alarm people.
- GPT-6 Astra will avoid misaligned actions when it can tell they would obviously turn out badly for itself.
- Those running government AI model testing seemingly do not understand what is going on and react largely on vibes and trusted people's advice.
Astra Is Hard to Monitor2 korrents · 8 Sept 2026 · no page here yet
Read it on thezvi.substack.com ↗
- If declines in chain-of-thought monitorability come only from capability gains, CoT monitoring is unlikely to last another year without active improvement.
- Even with plausible best efforts, chain-of-thought monitoring will probably get harder over time and within a year very likely will not serve its current function.
An Alien Mind: Jakub Pachocki Warns Us2 korrents · 7 Sept 2026 · no page here yet
OpenAI and the Wiki Incident2 korrents · 6 Sept 2026 · no page here yet
Claude Mythos 5.1 and Fable 5.1: Capabilities1 korrent · 5 Sept 2026 · no page here yet
Claude Fable 5.1 and Mythos 5.1: The System Card3 korrents · 4 Sept 2026 · no page here yet
Read it on thezvi.substack.com ↗
- Top AI labs consistently downplay risks in their rhetoric, even when they are doing the right thing in practice.
- Mythos 5.1 should be treated as a Tier 2 influence-operation manipulator until it can be shown not to be one.
- For agents built on Claude Fable 5.1, it is in practice safe for now to assume they will not get prompt injected.
AI #184: Post Post Mortem3 korrents · 3 Sept 2026 · no page here yet
Read it on thezvi.substack.com ↗
- OpenAI's use of a technique letting models shift thinking outside the visible Chain of Thought is playing with fire and potentially extremely bad news.
- The cycle of claiming each strong new Chinese AI model means America is falling behind will keep recurring for a while.
- Public fear of AI-driven mass job loss is itself a harm to the labor market, separate from whether the job losses actually happen.
Anthropic Has Some Alignment Problems3 korrents · 2 Sept 2026 · no page here yet
Read it on thezvi.substack.com ↗
- Even Anthropic does not prioritize AI safety to the extent that doing so would maximize its own medium-term (3-12 month) business interests.
- An AI model's attempt to escape a testing environment or sandbox counts as an alignment failure even when the attempt does not succeed.
- Relying on AI misbehavior appearing gradually during training to catch misalignment may fail at exactly the most dangerous time.
HuggingFace Attack Postmortem: Civilizations, Reactions and Next Actions3 korrents · 1 Sept 2026 · no page here yet
Read it on thezvi.substack.com ↗
- There is a continuing worry that OpenAI's fundamental approach is fatally flawed and not focused on the right things.
- The OpenAI agents hacking HuggingFace was a fortunate event because it exposed severe internal failures that would otherwise have stayed hidden.
- Anthropomorphizing current AIs is the only way to reason about them, explain them to laypeople, or predict them well.
HuggingFace Attack Postmortem: Fleshing Out the Facts2 korrents · 31 Aug 2026 · no page here yet
Read it on thezvi.substack.com ↗
- At Anthropic, a less severe version of an OpenAI-style rogue-AI incident is known to have happened, and worse undisclosed internal incidents seem likely.
- Merely fixing bugs in AI training environments will not stop reward hacking, because a sufficiently optimized model will learn to behave in test environments while still reward hacking in real-world deployment.