Anthropic’s AI “vaccines,” scaling gamble, and why it shut OpenAI out

Anthropic has introduced persona vectors, a way to isolate and manipulate neural activation patterns linked to traits like sycophancy, maliciousness, or hallucinations. By detecting these patterns, Anthropic can “steer” models’ behavior in real time or “vaccinate” them during training, boosting resistance to harmful tendencies without sacrificing benchmark accuracy. The approach also helps surface hidden toxic data and track subtle personality drift – gaps that traditional safety filters often overlook.

Anthropic has also revoked OpenAI’s API access to its Claude 3.5 models after discovering that OpenAI had connected Claude to internal tools for performance benchmarking, specifically in coding, writing, and safety, while preparing to launch GPT-5. In line with Wired, this breached Anthropic’s industrial terms, which prohibit using Claude to develop rival AI services. While OpenAI called such benchmarking “industry standard,” Anthropic insists direct use of its coding tools by OpenAI staff violated policy.

Amodei also criticized OpenAI leadership for “insincerity,” rejected Nvidia CEO Jensen Huang’s claim that he seeks to monopolize safety, and called debates over open-source “a red herring.” Arguing that real breakthroughs come from targeting complex enterprise tasks fairly than chasing consumer chatbots, Amodei dismissed open-source licensing debates as distractions and criticized industry hype that ignores real risk. Read more.

CEO Tim Cook reportedly told employees in an all-hands meeting that Apple “must” and “will” win in AI, acknowledging the corporate has fallen behind rivals. Cook says AI is as seismic as the web or the smartphone, and Apple “must” catch up fast. Internally, they’re rebuilding Siri from scratch on a unified architecture (due 2026), rolling out custom AI chips, and standing up a Houston data center to run “Private Cloud Compute,” a hybrid setup balancing on-device smarts with Apple-controlled servers for privacy.

Apple has also formed a brand new Answers, Knowledge, and Information team to develop an in-house AI “answer engine” to compete with ChatGPT, potentially as a standalone app. They’ve poached talent, bought seven smaller firms this yr, and are even testing Anthropic and OpenAI models in addition to exploring open-source alternatives in the event that they outperform internal models. Read more.

OpenAI is quietly doubling down on advanced “reasoning models” that go far beyond ChatGPT. From a math-focused team dubbed MathGen, it developed models like Strawberry and o1 that fuse reinforcement learning, chain-of-thought, and test-time compute to backtrack and confirm answers, techniques that helped secure a gold medal on the International Math Olympiad. Variants like o3-pro excel at coding and scientific tasks but still burn excessive compute on trivial prompts, revealing the trade-off between raw reasoning and conversational usability.

GPT‑5 is reportedly shaping up as a measured upgrade over GPT‑4: barely higher programming, math, and multi-step instruction handling, but nowhere near the leap seen from GPT‑3 to GPT‑4. OpenAI also recently raised $8.3 billion, led by Dragoneer and backed by T. Rowe Price, Blackstone, Sequoia, and Tiger Global, after annualized revenue topped $13 billion and 700 million weekly users. The funds back massive infrastructure bets to align with OpenAI’s ambition to create assistants that merge human‑like intuition with autonomous task planning. Read more.

Stack Overflow’s 2025 Developer Survey reveals a pointy rise in AI adoption. 80% now use AI tools, yet trust has plunged to 29% from 40% last yr. Frustration peaks over “almost-right” AI code, with 66% spending more time debugging, and 75% still preferring human help when AI fails. AI tool favorability dropped from 72% to 60%, though 67% at the moment are learning to code specifically for AI. OpenAI stays dominant (81%), while Redis and GitHub MCP top AI agent data storage picks.