cat briefings/2026-09-04.md

Daily AI briefing

One day's edition, kept at its own URL. The current briefing is published each morning.

Fri, 4 Sept 2026

IFM's Apache-2.0 K2 Horizon fleet, Jane Street's $13 billion compute contract with Crusoe, and the Justice Department siding with OpenAI on training as fair use.

generated · sources linked on every line

Africa & low-resource ML

  • Askya Investment Partners opened applications for the Askya AI Growth Platform, a six-week, zero-equity programme with up to $200,000 in investment for Africa-focused AI startups from pre-seed to Series A that have a live product and paying customers, with applications closing 30 September and the cohort running 28 October to 4 December. TechCabal
  • US cybersecurity firm Cybastion plans a $75 million AI data centre in Douala, Cameroon, paired with its own dedicated power plant to work around unreliable grid supply, agreed at the first US–Cameroon Bilateral Economic and Commercial Dialogue. Broadcast Media Africa
  • Skills for Progress Africa launched in Accra with an inaugural cohort of 10 fellows trained on the Festo LX platform using AI-supported industrial simulation, with ELAF Industrial Automation as founding technical partner and a stated target of 10,000 fellows. MyJoyOnline
  • FATE Foundation and AIMS launched the Google.org-backed Africa AI Upskilling Programme at Ghana Communication Technology University in Accra, training university students and educators in applied AI. Ghana Communication Technology University

AI for science

  • A machine-learning and statistical-physics pipeline that fingerprints the protein segments controlling attachment at an oil–water interface screened millions of candidates down to nearly 800 plant proteins usable as emulsifiers, with pea and potato proteins confirmed in lab tests. Phys.org
  • Nitesh Chawla and Olaf Wiest argue in Chemical Reviews that organic chemistry's small datasets, 3D structure and mechanistic constraints force AI systems toward interpretability and calibrated uncertainty rather than pure scaling, making chemistry a testbed for explainable models. Tech Xplore

Free offerings

  • The Institute of Foundation Models released K2 Horizon under Apache 2.0, led by a 375B-parameter sparse mixture-of-experts activating 23B per token with a 512K context, scoring 87.3 on GPQA Diamond and 42.6 on SWE-Bench Pro, alongside 32B, 7B and small GGUF builds, with training data and code to follow. new Institute of Foundation Models

Building on frontier models

  • OpenAI connected ChatGPT Health to Epic, letting authorised clinicians pull read-only patient context — notes, labs, medications, specialist documentation — into a ChatGPT conversation or run ChatGPT inside the Epic chart, with UCSF Health as pilot partner and nine public health data sources added. TechCrunch
  • CrowdStrike and Nvidia launched SafeMind, which pairs a Nemotron-based offensive model (Red Tempest, trained partly on 15 years of incident-response data) against a defensive one (Blue Solano) inside a digital twin of a customer environment, iterating until attack paths are closed. Nvidia

Cost & hardware

  • Jane Street signed a roughly $13 billion, five-year contract for GPU clusters and AI infrastructure on Crusoe's cloud, the largest cloud deal Crusoe has closed and a rare case of a trading firm buying frontier-scale compute directly. Data Center Dynamics
  • On the back of that contract, Crusoe is reportedly raising $3 billion at a $30 billion valuation. TechCrunch
  • Bitdeer bought 200 acres in Milam County, Texas, to expand its AI and HPC data centre footprint. Data Center Dynamics

Evals & guardrails

  • The US Justice Department filed a statement of interest urging Judge Sidney Stein to rule for OpenAI and Microsoft in the New York Times case, calling model training "extraordinarily transformative" fair use, the federal government's first formal position in the AI copyright litigation. The Washington Post
  • The Texas sheriff's office that searched Flock's nationwide camera network for a woman who self-administered an abortion used Axon's Draft One to draft the report on it, so the record cannot show which language came from the officer and which from the model. 404 Media

Algorithmic advances

  • Hardware-Aware FP4 FlashAttention-4 maps attention onto Blackwell's 4-bit tensor cores, reaching up to 2.13× BF16 forward throughput on a GB200 and accelerating a full single-GPU 8B-parameter update by up to 1.14× using FP8 gradient operands on the causal path. arXiv
  • Running on-policy distillation and then RLVR as separate sequential stages consistently beats pure distillation, pure RLVR and every joint variant, because distillation expands coverage of teacher-supported solutions while RL sharpens within that support and mixing the two signals causes interference. arXiv
  • LeanGRPO removes the redundant gradient-tracked recomputation of rollout timesteps in diffusion RL, giving up to 1.83× end-to-end speedup on FlowGRPO and DanceGRPO with FLUX.1-dev and Wan without changing the optimisation objective. arXiv
  • DRACO generates rubrics during training and redistributes rubric-level judgement into per-step advantages for GRPO, gaining 15.9 points over the base model and 5.3 over sparse-reward GRPO on AppWorld. arXiv