OpenAI’s chief scientist has thoughts on how to limit AI development
To establish a speed limit on AI development, you need to agree on how to measure speed.
To establish a speed limit on AI development, you need to agree on how to measure speed.
OpenAI Chief Scientist Jakub Pachocki has disclosed internal log-perplexity figures for successive models from GPT-3.5 onward and proposed that frontier labs use the metric as the basis for common industry safety standards. He identifies a 0.68–0.7 range as a known-unknown threshold where further transformative capabilities are expected but cannot yet be specified.
A quantitative analysis using Anthropic as a case study estimates that by 2028, about 87% of the electricity for its contracted U.S. data centers will come from fossil fuels, with renewables near zero. The report links AI compute contracts to actual power sources and quantifiable emissions, raising structural energy and carbon concerns.
MLCommons has released the MLPerf Training v6.1 results, adding recommendation and large-scale post-training benchmarks while highlighting growing hardware diversity and broader industry participation.
MLPerf Training v6.1 expands the benchmark suite and showcases larger systems submitted by members, adding two new benchmarks, debuting Google Ironwood and NVIDIA Rubin, and reaching a maximum job size of 10,240 GPUs.
MLCommons has released MLPerf Endpoints v1.0, a benchmark that measures full-stack AI performance at real-time API endpoints under production-like workloads. It evaluates complete systems using real customer workloads and a Pareto frontier rather than single peak scores, providing buyers and providers with practical performance insights.
On this pageOverviewMeasurement Setup1. Prefill CP and Decode DP2. FP8 DeepGEMM for the Prefill Indexer3. Breakable CUDA Graphs for Prefill4. Tensor-Parallel MoE5. Larger Indexer Score Tiles6. Fuse Pr
In today’s Smoke evaluation, Grok 4’s main leaderboard score fell from 83.51 to 66.35, a drop of 17.2 points, while the Material Constraints dimension plunged 24.8 points in one day. The decline is more consistent with question-sampling volatility than systematic degradation, though the integrity rating change and the low absolute Material Constraints score warrant caution.
In today's Smoke evaluation, Claude Sonnet 4.6's Material Constraints score fell by 24.8 points while its Code Execution score rose by 50 points, lifting the overall main leaderboard score from 61.56 to 77.90.
The 2026-11-23 YZ Index Smoke quick test covered 14 models, with Claude Opus 4.7 ranking first for the day at 78.72. These single-day results are monitoring signals rather than long-term conclusions.
Journalist Paris Marx unpacks his new book "Hyperscale."
Major US financial institutions are banning third-party bots from logging into your accounts—and warning that any costly bot mistakes are strictly your problem.
A researcher at NYU and a researcher at Anthropic released an AI-assisted proof on Mastodon claiming that the Navier-Stokes equations can exhibit finite-time blowup in the absence of external forces, and the same researcher also announced a claim regarding the Riemann Hypothesis the same day.
Mathematician Levent Alpöge released a 17-million-line Lean proof library for the Riemann Hypothesis built with Claude assistance, and the Lean and Mathlib communities independently compiled and checked it. Multiple independent tools and kernels confirmed that the proof matches Mathlib’s RiemannHypothesis statement and is valid.
Google has cut its 2027 Broadcom-designed TPU 8i shipment plan by about 1 million units, roughly 20%, while increasing orders for MediaTek-designed TPU 8t. The shift, shaped by power and capacity constraints, signals that training workloads now take priority over inference in Google’s AI chip allocation.
Amazon is in talks to acquire Israeli AI startup Decart for $7.5 billion in cash, equity, and earnout provisions, aiming to reduce AWS's dependence on NVIDIA GPUs through real-time inference optimization across multiple hardware platforms.
We got two very different pictures of Nvidia — the company’s latest earnings report vs. the tech executives loudly calling for Jensen Huang’s return.
Signal is limiting users' ability to take screenshots in both its Android and iOS apps.
Anthropic's record $81.3 billion IPO failed to reassure investors and instead coincided with a sharp selloff that erased the Nasdaq's 2026 gains, pushed the S&P 500 into correction, and triggered a broad repricing of AI stocks. The drop reflected three simultaneous pressures: high interest rates, Nvidia's regulatory crisis, and export controls.
Google has released Gemini 4 Neon, positioned as the efficiency-depth balance point in the Gemini 4 family, with roughly twice the output speed of the flagship Carbon. However, its current pricing is higher than Carbon's promotional price, creating a temporary price inversion until Carbon's price increases in December.