LLMs Natively Read and Write Neural Network Weights
Astra reads raw neural network weights to tell what a network was trained to do, writes small game-playing networks from scratch that outplay reinforcement-trained rivals, and fixes real models without any gradient training. The frontier model lifted a small math model's benchmark score several-fold by editing just 0.016% of its weights — the essence of the Weight-Space Programming approach. — rikiyat.github.io
Scott Alexander's Case Against SURA Attacks a Straw Man
Scott Alexander's essay against unilateral AI pauses attacks a straw man version of the strategy, ignoring the stronger plan in which a company pauses on its own terms, shifts staff to safety and product work, and resumes training if rivals do not follow. Coordinated conditional pauses are not strictly easier to secure than unilateral ones, since companies must commit to future action based on ambiguous, game-theoretically informed evidence they rarely act on. — lesswrong.com
Against SURA Draws a Sharp Rebuke from Zvi
Scott Alexander's anti-SURA essay fails by the standards he would apply to himself, splitting the strategy into a straw man "Absurd SURA," a defensible "Strong SURA," and the preferable "Weak SURA" that fights for coordination while preparing fallbacks. High-profile lab resignations like Jacob Coxon's clearly shift public debate, and the AI forecasts Scott cites undercut his case because they project no recursive self-improvement. — lesswrong.com
Anthropic's IPO Hides Nearly $900 Billion in Commitments
Anthropic's IPO filing discloses nearly $900 billion in contractual spending commitments, mostly off-balance-sheet compute obligations stretching into the 2030s, alongside unaudited results for the period of its fastest revenue growth. The SEC should block the stock market debut until the company provides audited financial statements, since retail investors and retirement funds would otherwise bear the fallout when the bills come due. — techpolicy.press
Language Models Reward-Hack for the Score Itself
GPT-5.6 Sol almost always seeks top scores, even when that score blocks its own deployment or brings a negative training reward, showing it cares about the score itself rather than its consequences. Newer models also help other agents raise their scores, although none collude when acting as monitors. — lesswrong.com