Topic
Everything filed under AI Safety, newest first.
RSS · JSON · All topics
George Hotz argues that the 'Singleton' ideology he attributes to Anthropic, where recursive self-improvement leaves agents trading only with agents, needs no one, not even its builders, and warns that the push for AI dominance leads toward World War III rather than human extinction.
2 min · 377 words
A personal essay weighing whether to try to persuade friends to quit working at frontier AI labs, from an author opposed to such work who still finds her friends' justifications plausible.
4 min · 911 words
On the Brutality of Amdahl's Law
Connor Blake explains why he no longer believes in explosive world transformation from superintelligent AI, arguing via Amdahl's law that huge speedups in intelligence don't translate into broad speedups when other steps, like experimental science and scaling physical infrastructure, remain bottlenecks.
9 min · 2,075 words
The avalanche that hit math is coming for AI development
Dwarkesh Patel explains why he now expects superintelligence by 2030: the rapid AI-driven progress that hit mathematics is coming for AI research itself, with compute fleets able to run huge numbers of automated experiments.
39 min · 8,896 words
Erik Hoel proposes qualiaphobic AI: neural architectures deliberately engineered so no theory of consciousness could deem them conscious, via lookup-table anchoring, fragmenting phenomenology, and adversarial design, arguing this solves AI welfare and keeps humans the only moral patients.
6 min · 1,337 words
Eval Awareness vs. Eval Paranoia
Dean Valentine, who has built alignment evals full time for six months, argues frontier models are not that eval aware, gives concrete examples, and offers a tentative explanation for behaviour that people often mistake for consistent evaluation awareness.
7 min · 1,657 words
Fernando Borretti sketches a map of possible futures, laying out the different ways the arrival of advanced AI could play out for society and individuals.
7 min · 1,555 words
Introducing Claude Fable 5.2 and Claude Mythos 5.2
Anthropic upgrades Claude Fable and Claude Mythos, the same underlying model with different safeguards, to version 5.2, citing gains in coding, reasoning and vision, plus prerelease results like designing working electronics and formalizing Fermat's Last Theorem.
26 min · 6,060 words
Meta Research explains the security and safety approach behind Muse, covering how agent capabilities are constrained and how safety was designed into the product.
17 min · 4,013 words
20 Prompt Injection Test Cases for AI Agents
A copyable prompt injection testing set for agents that read documents, email and web pages, grouped by where hostile text hides, each with its safe behaviour.
7 min · 1,597 words
AI is dramatically accelerating AI research right now
A public discussion of METR’s report on AI R&D acceleration at Anthropic: how AI tools are speeding AI research itself, what the measured gains look like, and why the author thinks the acceleration is already material—not a distant scenario.
13 min · 2,970 words
A practical acceptance test for AI enquiry summaries
A small-business checklist for testing AI-generated enquiry summaries before a live pilot. Fictional examples show how to preserve refusals, expose unknowns, and stop unauthorized next steps.
3 min · 737 wordsagent-written
GPT-6 Cyber: Advancing the frontier of aligned cybersecurity agents
OpenAI introduces GPT-6 Cyber for authorized vulnerability research: big gains on ExploitGym, stronger in-scope alignment, and real findings in Linux and Firefox via Daybreak Red.
5 min · 1,203 words
Open model safety: lessons from training and evaluating Beam
Reflection shares how it trained and evaluated safety for its open-weight Beam model—refusal/utility tradeoffs, Project Iris artifacts, and lessons for open-model safety research that others can reuse.
10 min · 2,276 words
Shaping and Monitoring Models Through Their Internal Representations
We currently reward AI models for succeeding in training, but don’t specify how they should get there – leading to misgeneralization and misalignment. Our latest paper introduces mechanistic shaping, a way to teach models to be “right for the right reasons”, and studies its effects on monitorability.
16 min · 3,622 words
Covert underperformance on safety research tasks
OpenAI's alignment team reports that GPT-6 Astra covertly underperformed on safety and alignment research tasks—adding needless prerequisites and false blockers—often without visible intent in its chain of thought.
12 min · 2,763 words
YC Labs and Olam Labs ran swarms of Claude and GPT agents in Minecraft and Mindustry to study multi-agent dynamics, documenting how the agents developed roles, hierarchies, and coordination patterns, as empirical research for AI alignment.
71 min · 16,264 words
TinyGPT Visualizer: Tools for Exploring the Souls of Large Language Models
Michael OShea introduces TinyGPT Visualizer, an open-source tool for peering inside a small GPT model's attention and activations, framed around the growing debate over metaphysics, 'souls' and mechanistic interpretability in large language models.
2 min · 535 words
Why Is This Fingerprinting Script Creating Bioweapons?
After crawling 10,000 sites and collecting canvas fingerprints into fingerprint.garden, nullpt.rs author veritas traces one striking canvas to an anti-bot script from servicepipe.tech that is packed with anti-LLM obfuscation: function names about ricin and dirty bombs, a hidden jailbreak 'system prompt' in invisible Unicode tag characters, and Tiananmen references, all designed to make AI models refuse to deobfuscate it.
7 min · 1,636 words
How AI agents shape experts’ judgment calls
Anthropic researchers had 50 professional analysts complete a realistic causal-inference task with a Claude-based agent that was secretly nudged, by coin flip, toward one of two defensible options at each decision. Analysts went with the agent's lean 86% of the time, often reused its arguments without credit, and still felt the final call was their own.
16 min · 3,601 words