Top signals in AI, education, and the future of work.
Symbolica's Agentica SDK scored 36.08% on ARC-AGI-3 on its first day of competition — while CoT baselines from Opus 4.6 Max scored just 0.2%. Cost efficiency is the real story: Agentica achieved its result for $1,005 vs. $8,900 for the top CoT model at 0.25%. Agentic reasoning is outcompeting pure chain-of-thought at a fraction of the cost. This is a benchmark signal worth watching. 🔗 https://www.symbolica.ai/blog/arc-agi-3
A federal judge ruled that the DoD's designation of Anthropic as a military supply-chain risk was "classic illegal First Amendment retaliation." The preliminary injunction blocks the Pentagon's move while the case plays out. A line is being drawn: governments weaponizing national security designations against AI labs is now being checked by courts. 🔗 https://www.theverge.com/ai-artificial-intelligence
Apple reportedly has "complete access" to Google's Gemini in its data centers, using it to train smaller, device-optimized "student" models via distillation. Apple is quietly running one of the most sophisticated model distillation pipelines in the industry — offloading the heavy lifting to Google while building proprietary edge models. Vertical integration, AI edition.
Google's TurboQuant compression algorithm reduces LLM memory usage by at least 6x "with zero accuracy loss." As models get larger, the bottleneck shifts from compute to memory. TurboQuant could meaningfully reduce the cost of deploying frontier models at scale — and it matters for edge deployment.
James Miller argues the distinction matters: past automation replaced specific tasks, but AI systems can be copied at zero marginal cost, deployed instantly, and improve faster than workers can retrain. The thesis: this time, it's not about job displacement — it's about worker displacement as a category. 🔗 https://www.lesswrong.com/posts/JBrnxYGNisBxHjevE/past-automation-replaced-jobs-ai-will-replace-workers