JS Denain

Commentary by JS Denain

From the commentary rail — every link leaves the site for the original piece.

• 7 July 2026 · Epoch AIThe missing half of AI futurism debatesWhy we should think a little harder about what it takes to build a Dyson Sphere
• 24 June 2026 · Epoch AIWhat we learned from 1,604 Chinese AI job postingsInferring Chinese AI labs’ strategies from their job descriptions
• 17 June 2026 · Epoch AIToward an O*NET for AI R&DProposing a new way to track AI research automation
• 11 June 2026 · Epoch AIAre Mythos’ cyber capabilities overhyped?Compiling all the public evidence on Mythos Preview’s cyber abilities
• 26 March 2026 · Epoch AIWhat do frontier AI companies' job postings reveal about their plans?A fast increase in go-to-market roles, and hints about upcoming products
• 24 March 2026 · Epoch AIFinal training runs account for a minority of R&D compute spendingNew evidence following the MiniMax and Z.ai IPOs
• 16 February 2026 · Epoch AIHow persistent is the inference cost burden?RL scaling might be better than it looks, and inference costs to reach a capability level fall fast
• 12 January 2026 · Epoch AIAn FAQ on Reinforcement Learning EnvironmentsWe interviewed 18 people across RL environment startups, neolabs, and frontier labs about the state of the field and where it's headed.
• 23 December 2025 · Epoch AIWhy benchmarking is hardRunning benchmarks involves many moving parts, each of which can influence the final score. The two most impactful components are scaffolds and API providers.
• 20 December 2025 · Epoch AIThe changing drivers of LLM adoptionPublic data as well as our original polling suggest LLM adoption is roughly on trend, but the underlying drivers are shifting.
• 17 December 2025 · Epoch AIIs almost everyone wrong about America’s AI power problem?Why power is less of a bottleneck than you think.
• 4 October 2025 · Epoch AIHow many digital workers could OpenAI deploy?OpenAI has the inference compute to deploy millions of digital workers, but only on a narrow set of tasks – for now.
• 26 September 2025 · Epoch AIWhy GPT-5 used less training compute than GPT-4.5 (but GPT-6 probably won’t)OpenAI focused on scaling post-training on a smaller model
• 19 September 2025 · Epoch AIThe huge potential implications of long-context inferenceContinual learning, scaling RL, and research feedback loops
• 23 August 2025 · Epoch AIWhy future AI agents will be trained to work togetherMany multi-agent setups are based on fancy prompts, but this is unlikely to persist
• 13 June 2025 · Epoch AIWhat does SWE-bench Verified actually measure?It is one of the best tests of AI coding, but limited by its focus on simple bug fixes in familiar repositories
• 8 June 2025 · Epoch AIBeyond benchmark scores: Analyzing o3-mini’s mathematical reasoningWhy o3-mini is a "vibes-based inductive reasoner"
• 28 March 2025 · Epoch AIThe real reason AI benchmarks haven’t reflected economic impactsThe real reason that AI benchmarks haven’t reflected real-world impacts historically is that they weren’t optimized for this, not because of fundamental limitations – but this might be changing.