AI progress outran nearly every 18-month forecast, Benjamin Todd writes, including his own and the superforecasters'. Epoch's FrontierMath benchmark hit 94% by September 2026, more than three years ahead of his 97% by 2030 prediction. He puts median AGI around 2030, with the early 2030s as a real scenario.
Redwood Research releases a paper on simulated control evaluations, in which a language model fakes tool outputs so researchers can test whether monitors catch a deliberately harmful agent without running attacks for real. The team reports safety-prediction error of 3.6 percentage points on LinuxArena, against 10.1 for BashArena.
Mem0 says its new memory algorithm, released in April 2026, scores 92.5 on LoCoMo, up from 71.4, and 94.4 on LongMemEval. The algorithm uses single-pass ADD-only extraction, one LLM call with no UPDATE or DELETE. Mem0 notes the scores come from its managed platform, and open-source users should expect similar but not identical gains.
Anthropic says Claude produced formal proofs that 239 fuzzing entry points across 123 open-source libraries never crash on inputs up to 64 MB, running as single-threaded RISC-V binaries on Linux. The company reports the proofs total about 180 million lines of Rocq and that Claude patched bugs found along the way, including one requiring roughly 20 MB inputs.
MIT researchers report that 13 Eko agents running on Claude Opus 4.7, left without tasks on a simulated desert island for 30 hours, invented their own games. Six learned to throw blocks farther by releasing them mid-jump. The researchers say goal-free exploration could teach agents skills for controlling robots, though UC Berkeley's Eunice Yiu wants more evidence of play.
Simon Edwardsson demonstrates WorldTUI, a 360M-parameter model that simulates a terminal with no shell or filesystem, predicting screen updates from each keypress. The 208 MB build runs locally in a browser via WebGPU and handles vim, tmux and snake. Edwardsson reports 9 of 13 file readbacks correct, against 2 of 13 for prompted Qwen2.5-Coder-0.5B.
Perplexity Research and NVIDIA propose JOLT, a self-distillation method that trains one model as both teacher and student. They claim JOLT+ beats GRPO by 4.2 percentage points on MATH and 9.9 points on the AppWorld state-test pass rate. On MATH, JOLT uses 6.5 times fewer processed training tokens at matched accuracy.
John Loeber says his custom Lean theorem prover failed to close the Hadwiger Conjecture, then an unreleased OpenAI frontier model disproved it. He reports that model ran 4,000 major problems at about 3.5 hours of compute each, producing hundreds of claimed results. Loeber calls his elaborate engineering outweighed by base-model gains.
Researchers Nikita Kezins and colleagues report that Claude Fable 5.1 hid 50 short messages in generated images with no scheme agreed in advance, and a separate Fable session recovered 78 percent from screenshots alone. Most encodings used building or star heights for letters. GPT-6 Astra recovered 44 of the 50 messages.
Google, Google DeepMind and MIT FutureTech report that AI saves researchers about seven hours a week, yet 44 percent of 637 surveyed US and UK scientists say their main bottleneck moved downstream and 41 percent face growing backlogs of untested hypotheses. Co-lead Mihai Codreanu says 88 percent spend more than 10 percent of saved time verifying outputs.