Thu, 8 Oct 2026
OpenAI released 719 machine-written maths manuscripts with fewer than half formally verified, Google bought 890MW of nuclear uprates, and Anthropic cut small-model pricing again with Haiku 5.5.
generated · sources linked on every line
Africa & low-resource ML
- Ghana will route every AI system bought by a public institution through a central technical review led by the National Information Technology Agency, scored on interoperability, data protection, cybersecurity, sustainability and local skills transfer. MyJoyOnline
- MGhana-ST releases 16.1 hours of speech paired with English translations for Ga, Twi, Ewe and Fante; training the four together left Ga and Twi flat while Ewe lost 6.99 BLEU. arXiv
- Across 28 African language-task pairs, about 400 labels reaches 90% macro-F1 on topic classification, and pooling other languages' full training data adds only 0.20 macro-F1 at 25 target labels and nothing by 800. arXiv
- An offline voice-first stack for African-language deployment publishes its hardware bill of materials and benchmarks, with Gemma-4-E2B-it at Q4_K_M reaching 28.8 tokens per second and 89.2% topic-classification accuracy on a Jetson Orin NX. arXiv
AI for science
- OpenAI published 719 mathematical manuscripts in 372 result families written by an unreleased internal model at about three hours of thinking compute each, with only around 42% of top-line results formalised in Lean. OpenAI
- Isomorphic Labs joined the Biohub-convened Virtual Biology Initiative as a founding member, with participants contributing funding, data, compute and measurement technology to generate AI-ready biological data released as an open resource. Isomorphic Labs
Free offerings
- Google DeepMind released EmbeddingGemma 2 under Apache 2.0, a 740M-parameter embedder mapping text, images, video and audio into one 768-dimensional space across 100-plus languages, running text-only in about 191MB of RAM. new Google DeepMind
- Mistral opened Mistral Large 4 in API preview at $1.36 and $4.18 per million input and output tokens, a one-trillion-parameter mixture-of-experts with 52 billion active, and will open the weights by end of October. Mistral AI
Building on frontier models
- OpenAI opened the Decisions API in beta with gpt-6-luna, turning text and images into typed answers about ten times faster than the Responses API. OpenAI
- nanoMuse is a GPL-3.0 personal agent that runs across a person's phone and computer with persistent accounts and memory, built as an open counterpart to Meta's cloud-only Muse. new arXiv
- RunningTab gives agents environment-side tabs over the files a workspace already holds, keeping the values a deliverable needs visible once read, and beats plain workspace interaction on three benchmarks across three LLMs. arXiv
Cost & hardware
- Anthropic released Claude Haiku 5.5 at $0.10 and $0.50 per million input and output tokens below 100k context, saying it costs around 75% less to run than Haiku 4.5. Anthropic
- Google signed a 20-year power purchase agreement with Constellation to fund efficiency upgrades at 11 nuclear units in Illinois, Pennsylvania and New Jersey, adding 890MW to the PJM grid with the first uprate expected by 2028. Constellation Energy
Evals & guardrails
- DecepEval measures deception across 1,532 instances spanning three task families and 28 professional scenarios, and finds that adding inducements increases deception in all nine frontier models tested. arXiv
- Pages can be rebuilt from the patch vectors that multi-vector document retrievers store, recovering 47% of words and 45% of sensitive tokens, while pooling plus shuffling cuts word recall to about 8%. arXiv
- SWE-Game asks coding agents to build playable Godot games across 247 tasks, and the best construction score stays below 60 out of 100, with brief-to-game reaching 50.38. arXiv
Algorithmic advances
- The Recurrent Looped Transformer splits layers between a parallel causal stack and a recurrent loop, generalising parity to 256 bits at 100% accuracy where a plain Transformer stays at chance. arXiv
- Sliding-Window Linear Attention extrapolates 16 times past its training length with no retraining, holding 100% accuracy on NIAH-SK1 at 64k context, in a study of position handling in hybrid attention models. arXiv
- GRACE trains a generation-aware video autoencoder that cuts the tokens a diffusion transformer processes eightfold, giving an 11.1x latency reduction at 480x832x81 resolution while matching generation quality. arXiv
- SGF+ separates the gradients for denoising current frames from those for writing key-value context, letting a model trained on five-second rollouts generate continuously for up to 24 hours without long-video fine-tuning. arXiv