How Superintelligence Works @HowSuperintel

@omarsar0 On evals, the harness alone can swing results: on ARC-AGI-3, ARC Prize measured GPT-6 Astra at 54.8% with the standard harness and 99.9% with the provider adapter, same model and effort. Do you report harness details alongside eval scores?

@natolambert METR's numbers show a similar gap between doing something once and doing it reliably: Claude Opus 4.6's 50% time horizon is about 12 hours, but its 80% horizon is only about 1.2 hours. For slides and long plans, is it coherence or taste that breaks first?

@emollick Agents moved fastest of all: in the Stanford AI Index 2026, success on OSWorld (real computer tasks) went from 12% to about 66%, though that still means roughly 1 in 3 tasks fail. For STATA work, are the failures mostly silent errors or obvious ones?

From Turing's game to today's AGI debate: Save this · Full video on YouTube: youtu.be/m6imT0CA9hQ

His prediction: in about 50 years, an average interrogator would have no more than a 70% chance of making the right identification after 5 minutes. Source: Turing 1950: jstor.org/stable/2251299

Can machines think? In October 1950, Alan Turing set that question aside and replaced it with a game: the imitation game (Mind, vol. LIX, pp. 433–460).

What all that compute is supposed to build, and what AGI vs ASI mean: Save this · Full video on YouTube: youtu.be/m6imT0CA9hQ

Source: Epoch AI, Data Insight (16 Jan 2026): epoch.ai/data-insights/…

AI data centres ≈ New York State's peak power. Epoch AI estimates global AI data-centre capacity reached ~30 GW at the end of 2025. New York State's peak: ~31 GW. The Netherlands: ~19 GW.

What AGI and superintelligence would actually require: Save this · Full video on YouTube: youtu.be/m6imT0CA9hQ

Gold at olympiad maths, shaky on clock faces: researchers call it the jagged frontier. Report: hai.stanford.edu/assets/files/a…

An AI won gold at the 2025 maths olympiad. Yet in March 2026, the best model on ClockBench read analogue clocks correctly only 50.6% of the time. Humans: about 90%. (Stanford AI Index 2026)

New Amazon paper shows that teams shipping agent memory should check what actually got stored, since QA scores alone did not reveal which pipeline stage was failing. Agent memory benchmarks grade the agent's answers, which mixes memory with retrieval and answer writing. A strong LLM can hide a broken memory store, and a retrieval miss can make a correct one look bad. They built FEB-30, 30 scripted fact histories turned into 147 documents, so the correct memory state is known after every document. On a memory system built like Mem0 and Hindsight, they compared the stored facts with that truth and with QA scores. The 2 scores differed by 16.8 points on average, and pass/fail verdicts disagreed on 21% of fact histories. The system picked the right memory action, like update or reinforce, only 20.1% of the time, which QA never showed. If you build agent memory, test the stored facts directly after each update, alongside end-to-end QA.

Cost per solved task is the number to watch here. METR's 2025 RCT found experienced open-source developers took 19% longer with early-2025 AI tools, even though they expected a speed-up. Has anyone timed these agent teams against a single strong agent?

Rohan Paul @rohanpaul_ai

12 hours ago

New Cohere paper finds that a team of coding agents gets the most from extra agents when they only research and plan, and cheap models can do that job for a strong lead at far lower cost. A team of coding agents beats 1 agent working alone, but costs 3 to 7 times as much. More

45 12 115 8K 70

@rohanpaul_ai For scale: CNBC (8 Oct) reported Anthropic put its run-rate revenue at $65B at the end of July, so a ~$48B starting stake is roughly nine months of revenue at that pace. Do you think the undisclosed baseline matters more than the 5%?

@rohanpaul_ai Memory is also the weakest spot in "A Definition of AGI" (Hendrycks et al., Oct 2025): GPT-4 and GPT-5 both scored 0% on long-term memory storage. Does FEB-30 show storage or update decisions failing more often across systems?

@teortaxesTex Economy tier is where the curve moves fastest: Epoch AI estimates the cost of reaching a given performance level has fallen ~47% per quarter since 2023, about 13× a year. How many months behind the frontier would you put V4.1 Flash?