The Information Machine
The edition

Thursday November 19, 2026

This is one edition of my daily email.

One free email each morning, around 8 AM US Eastern.

New today

01
Concluded today

OpenAI's AI research acceleration report

  • OpenAI research on chain-of-thought controllability, released alongside its transparency report on AI use in research, found the property increases substantially during reinforcement learning even without explicit process supervision and can be distilled from a more controllable teacher model.
  • The transparency report documented the median researcher spending $2,400 per day on AI agents, with agent hours growing 11-fold while experiments grew only 68%; for GPT-6.2 Astra, Codex helped surface key research insights, which OpenAI describes as AI's first reported foothold in research taste.
The gist

OpenAI's data shows AI is materially accelerating parts of its own development, while the gap between agent hours and output growth raises questions about how much acceleration translates to capability gains. The first reported AI influence on research taste in model development marks a qualitative shift in how OpenAI characterizes its agents' role.

02
New

Google DeepMind's Gemini 4 Neon

  • Google DeepMind launched Gemini 4 Neon on November 18, a mid-tier model ranking 1st on the Artificial Analysis Coding Agent Index at 58.6 and scoring 51.26% on WebExploitBench v1.2 at half the cost of competing large models.
  • Simon Willison noted a pricing inversion: a temporary discount puts Carbon at $2/$10 per million tokens against Neon's $3/$15 until Carbon doubles to $4/$20 on December 1.
  • Researcher Buck Shlegeris separately flagged on November 12 that GDM's Argon safety report appears to describe intentional CoT training, contradicting GDM's stated policy of not training on CoT.
The gist

Neon enters a competitive segment with specific cost and benchmark claims against rival frontier models. The temporary pricing inversion, where the mid-tier model costs more than the tier above it, creates practical confusion for developers choosing between models before December 1.

03
New

OpenAI AI models applied to chip design

  • OpenAI announced that its AI models identified up to 14% improvement in power, performance, and area on AMD's flagship GPU IP in under five months, building on designs AMD engineers had spent years optimizing.
  • Greg Brockman described the gains as compressing a year's worth of generational chip improvements into five months, characterizing them as additive to human engineering work rather than a replacement for it.
  • OpenAI applied the same approach to Jalapeño, its first custom inference chip, with gains verified on real silicon, and said it is partnering with Synopsys to integrate its model capabilities with Synopsys's EDA tools for broader access by chip designers.
The gist

The results show AI contributing measurable improvements to chip designs that experienced engineers had already heavily refined. OpenAI applying the same approach to its own custom inference chip indicates the technique is being used in production hardware development.

04
New

DeepMind's Project Genie interactive world model

  • DeepMind released Genie 4 on November 18, 2026, adding real-time native audio that adapts dynamically to scenes, including ambient soundscapes and character dialogue, and mid-session editing via natural-language prompts to change weather, characters, and environments.
  • A new Project Genie agent lets users describe a vision conversationally while Google's multimodal models build visuals and custom background scores.
  • Genie 3 also received upgrades: session length extended from 60 seconds to 3 minutes, plus world chaining, 360-degree panorama image grounding, and real-time in-world events.
  • All outputs carry SynthID watermarking and pass safety screening; Genie 4 is available to Google AI Ultra subscribers.
The gist

Genie 4 is the first version of the model to include native audio and mid-session world editing, expanding what users can do within a single interactive session. The addition of a conversational creation agent lowers the input barrier for building interactive environments.

05
New

Genesis AI humanoid robot funding

  • Genesis AI closed a $500 million Series A at a $3.5 billion valuation to develop general-purpose humanoid robots.
  • The company captures dexterous hand skills from human workers wearing gloves that record touch and force data, feeds those into its GENE model, and evaluates training in Genesis World, a physics simulator running parallel virtual environments to vary scenes, objects, and conditions.
  • The company's stated goal is robots with "human level intelligence and capability" that "work reliably across entire workflows for years", though its flagship humanoid was not ready for a public demonstration at the time of the announcement.
The gist

A $500 million Series A at a $3.5 billion valuation is a substantial funding event for a humanoid robotics company still without a publicly demonstrable flagship robot. The approach of learning from instrumented human workers and validating in simulation before physical trials represents a specific technical bet on how to close the gap between robot capability and real-world dexterity.

06
New

Perplexity's multiagent mixed-model teams feature

  • Perplexity launched a Multiagent effort level on November 17 in its Computer and teammates products, where a lead agent splits tasks across a team of agents each running different frontier models.
  • As the agents work, they share findings with one another, and the lead agent combines the results into a single answer.
  • The feature is available to Max and Enterprise Max web users.
The gist

The feature applies coordinated multi-agent workflows to research tasks, with different frontier models contributing to a single answer. It is available only to Max and Enterprise Max tiers.

07
New

AI agent benchmarks for ML engineering

  • Singularity-Bench's authors stated the Carnegie Mellon ML engineering benchmark is a proxy for recursive self-improvement skills rather than a test of the full autonomous loop, while SWE-CoWork's authors identified code translation as the bottleneck, with more than half of Opus 5.5's failed checks occurring after the agent already had the requirements.
  • Six of eleven Singularity-Bench inference engines scored zero.
  • A fourth benchmark, FrontierChallenge, found frontier models completing only 20.6% of 97 scientific workflow tasks despite frequent false claims of completion.
The gist

These benchmarks move beyond isolated coding tasks to measure whether agents can handle full ML engineering workflows, including training pipelines and multi-step team coordination. The gap between passing individual checks and completing full tasks on SWE-CoWork, and the acknowledged scope limits of Singularity-Bench, illustrate how current evaluations constrain what conclusions can be drawn about agent capability.

Get the daily email

I send one email each morning with the stories that moved. If you would rather just read here, that works too.

One free email each morning, around 8 AM US Eastern.

See the latest edition · Privacy

Updates

08
Day 2

ByteDance Malaysia GPU cluster financing

  • Tencent president Martin Lau said the November 16 US licensing rule, which froze a $480 million loan for ByteDance's Malaysia GPU cluster, closes the offshore workaround Chinese labs had used to train flagship models on Blackwell chips rented abroad, and that training will slow with model releases possibly delayed.
  • Tencent CSO James Mitchell said inference and cloud workloads can run on domestic chips already.
  • Aolani Cloud said it is working with regulators and financing partners to maintain compliance.
The gist

The rule closes an offshore compute workaround that Chinese AI labs had been using to access advanced chips, with Tencent's president saying it will slow AI model training and may delay releases. The legal uncertainty it creates for private cross-border financing arrangements is broad enough that multiple banks are reconsidering participation in deals that have no direct US nexus.

09
Day 2

Perplexity's Frontier Research division launch

  • Synthetic Sciences, a Y Combinator startup, released FrontierML on November 19, a benchmark of 25 ML research problems in which agents receive 12 hours and H100 GPUs but never see the held-out evaluation objective and must each design their own proxy metric for self-guidance.
  • None of five tested frontier models met the strong-result threshold.
  • The release came days after Perplexity announced on November 16 a new research division, backed by more than 10,000 GPUs, to build systems capable of autonomous research.
The gist

Perplexity is committing significant compute toward systems that conduct research autonomously rather than assist humans doing so. The FrontierML results show that current frontier models, when deprived of a known evaluation objective, do not meet a strong-result bar on open-ended ML research problems.

10
Day 10

Hyperscaler Q3 earnings and AI supply crunch

  • Nvidia's full Q3 results landed at $115.8 billion in total revenue, up 103% year-over-year, with $130 billion Q4 guidance, while CoreWeave deployed Vera Rubin NVL72 racks at 200 kilowatts per rack and GPU rental prices for B200s reached $6.97 per hour.
  • At a Semafor conference, Mubadala CEO Khaldoon Al Mubarak acknowledged a coming 'winter' in the AI buildout but said 'winter doesn't mean minus 30 degrees.
  • It could be a nice pleasant winter,' while executives ruled out stopping the buildout altogether.
The gist

Nvidia's results confirm that AI hardware demand is extending well beyond hyperscalers into enterprise and neocloud customers, with the company's quarterly profit alone now exceeding its total revenue a year prior. Supply constraints at every level, from laminate and memory to fab equipment and cooling, mean the infrastructure cycle is bounded by physical production limits, not just capital availability.

11
Day 27

Nvidia Rubin memory cuts and HBM shortage

  • CXMT, China's largest DRAM maker, said on its earnings call that DRAM supply will remain tight through Q4 2026 and reported gross margins near 80% despite lacking EUV equipment.
  • Huawei plans to source HBM for its Ascend 960DT accelerator from Samsung and SK Hynix because CXMT's HBM3E production yields of 20-30%, roughly half SK Hynix's approximately 70%, produce samples that perform closer to the older HBM3 standard; unofficial HBM3E prices in China have reached $4,000 per stack.
The gist

HBM supply constraints are forcing spec reductions on both shipping and next-generation Nvidia GPUs, affecting the memory capacity hyperscalers can deploy. Leading HBM suppliers are booked through 2029, and Chinese domestic HBM production cannot approach competitive yield or performance levels, keeping global AI accelerator supply dependent on a narrow set of Korean memory makers.

12
Day 6

IFM, Runway, and NVIDIA world action models

  • Runway's Praxis-1, the Apache 2.0 world action model released November 11, is a 27B-parameter model that jointly predicts video and actions and supports LoRA fine-tuning for adaptation to custom robots, cameras, and tasks.
  • Its NVFP4 checkpoint also runs on NVIDIA RTX PRO 6000 Blackwell, DGX Spark, and DGX Station, with the checkpoint available on Hugging Face.
The gist

Both models are open-weights with permissive licensing, making physical-AI simulation and robot policy testing more accessible. The reproducibility gaps flagged in Praxis-1's evaluation reporting leave some of its performance claims difficult to verify independently.

What is moving now · All editions · All stories