DISCOVER THE FUTURE
AI is getting weird. Keep up.
Page 3
PSI released an AI-assisted proof of the Aaronson–Ambainis conjecture and public Lean code. It says a complete proof emerged on day three. Humans built the system and reviewed the work; computing costs remain undisclosed.
A Transluce and UC Berkeley study turns isolated training surprises into a large experimental collection. Some deliberately constructed datasets look routine but change behavior outside their apparent subject, and researchers struggle to predict the effects by reading them.
In MAMUBench's simulated group orders and bookings, personal-agent teams trailed a single coordinator. Blocking checkout until messages were read rescued more failed runs than appointing a leader.
BypassGym puts AI-written programs on Windows virtual machines running real security products. Several lightweight models succeeded, but a passing test does not mean the security software raised no alarm.
TestPrism shows why checking one reference solution can make generated tests look more reliable than they are.
WikiTrace links an AI’s condensed lessons back to the actions and feedback that produced them, giving the agent a way to question its memories before acting.
In a research guessing game, weaker models could select useful questions supplied by other models but struggled to invent them. More turns alone did not solve the problem.
Tree-Ising targets repetitive reasoning through weight edits.
New box-office reporting on Sanxingdui: Future Past shows how far generating a feature-length movie can be from attracting paying audiences.
Stanford and Genesis Molecular AI tested a difficult drug-discovery judgment: choosing the more realistic of two molecular poses. A hybrid system using AI-written features outperformed the observed expert results.
In simulation, an AI-controlled arm followed a bogus instruction displayed in its surroundings. Researchers tested staged safety checks across simulated and physical robot tasks.
A controller mistook waiting for progress. Successes fell from 47 to 16 out of 50, then recovered to 45 when pauses were hidden from memory.
Apple's AMP1 improvised tool use and succeeded in 104 of 108 physical trials, but successful tasks averaged 280 seconds.
Benign memories can push agents beyond the scope of earlier instructions, a study finds.
An AI-generated, AI-reviewed collection is correcting anatomical errors. Its latest update includes 252 regenerated stickers.
CloudSEK says an exposed Silent Lynx server contained Hermes agent records linked to the collection of more than 2,000 Tajik government files.
A Sereact/KIT study questions how much familiar robot tests reveal about generalization.
Google released Gemini 4 Hydrogen, the fast, low-cost member of its Gemini 4 family, on Nov. 12 at 1:00 p.m. EST and is rolling it out to all Google Search users worldwide and in the Gemini app. Google says it edges out its pricier Argon model on a third-party intelligence index, runs 2.5x faster than Gemini 3.8 Flash and cuts fabricated claims more than fivefold, though its own model card shows it still trails a rival on some coding and computer-use tests.
New developer tools expand access; supplier-site restocking remains in beta.
Users report that Anthropic's Claude has started refusing to help build Pokémon fan games and assets, even when the sprites are already on the user's own computer. Japanese outlet AUTOMATON says it extends a late-October Claude Code rule against recreating specific existing games, and that the rule text now spells out names, characters, art, music, levels and appearance. Anthropic has not publicly commented on the Pokémon refusals.