cat briefings/2026-10-14.md

Daily AI briefing

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

Wed, 14 Oct 2026

Gemini 4 Argon reached general availability, ASML's quarter showed AI equipment demand still climbing, and Ghanaian languages gained two substantial new speech resources.

generated · sources linked on every line

Africa & low-resource ML

  • Nufo releases the largest Ewe speech-recognition benchmark to date, 174 standardised hours, on which Whisper-medium reaches 37.98% word error rate and AfriHuBERT 39.70%. arXiv
  • Ghana Open AI, supported by Ghana NLP, published a 91.33-hour Twi (Akan) speech corpus under CC-BY-NC-4.0: 10,960 verified single-speaker clips with UTMOS and DNSMOS quality scores plus IPA and dual transcriptions. new Ghana Open AI
  • Back-translating monolingual text beats continued pretraining by 7.7 to 16.4 XCOMET-XXL points for low-resource translation, and with only 1,000 parallel pairs in Odia and Swahili it beats a model given ten times more parallel data. arXiv
  • Pawa AI released Mwamuzi-Data v0.2 under CC-BY-SA-4.0, a Swahili-first decision and guardrail training set of 29,350 rows spanning language ID, jailbreak and harm labels, and Tanzanian mobile-money scam SMS. new Pawa AI

AI for science

  • MFG turns molecular optimisation into a multi-turn feedback loop, lifting average similarity-constrained success across ten multi-objective tasks to 81.50% from 70.81% for the base GeLLM³O model. arXiv
  • An all-electron machine-learned orbital-free density functional for periodic CHNO crystals cuts held-out density error from 6.7% to 0.4% of the electron count, with 0.25 mHa per atom energy error. arXiv
  • A screen of 106,235 Heusler structures, combining machine-learned densities of states with 3,625 first-principles transport calculations, certifies 80 thermoelectric candidates, 61 of which have never been measured. arXiv
  • X-Ragent trains chest X-ray reasoning with tool-augmented reinforcement learning on 7.7K Rad-CoT samples, gaining 8.6 points on the fully held-out ChestAgentBench. arXiv

Free offerings

  • Liquid AI released LFM2.5-24B-A2B as open weights, a 24B mixture-of-experts with 2.45B active parameters and 131,072-token context that scores 48.09 on LiveClaw against 38.19 for Qwen3.6-35B-A3B, running at 184 tokens per second on an Apple M5 Max. new Liquid AI
  • Fish Audio published Drama 3 weights under a research licence: a 35B mixture-of-experts text-to-speech model with 3B active parameters, up to 16 speaker tags in one script, coverage of 80-plus languages including Yoruba, Swahili, Amharic and Shona, decoding 271 frames per second on one H200. new Fish Audio

Building on frontier models

  • Gemini 4 Argon reached general availability on 13 October with a unified 1,048,576-token input and output limit, configurable thinking levels and agentic video understanding, while its Cyber Permissive variant stays behind Fairwind allowlisting. Google Cloud
  • Agent-Reflex trains a Qwen3-8B agent to act on noisy MCP tool output and mid-task user corrections, scoring 58.6 across five long-horizon tool-use benchmarks and 29.1 on AgentNoiseBench. arXiv
  • ReTerm teaches terminal agents to recover from their own command errors, and ReTerm-9B reaches 34.8% Pass@1 on Terminal-Bench 2.0, 14.6 points above its Qwen3.5-9B base, with code, data and model released. arXiv

Cost & hardware

  • ASML reported Q3 net sales of €11.7 billion and net income of €5.0 billion, guiding to €14.2–15.2 billion for Q4 and €44–45 billion for the year, with CEO Christophe Fouquet saying customers are raising equipment forecasts on the size and duration of AI investment. ASML
  • Intel detailed its GPU D-Series, formerly Crescent Island: 1.3 petaflops of FP8, 1.5 TB/s bandwidth and up to 480GB of LPDDR5X in a 350-watt air-cooled PCIe card holding a 300-billion-parameter model, now sampling to customers. Intel

Evals & guardrails

  • Across 15,555 health conversations donated by users in India, Nigeria, Brazil and Pakistan, roughly one in five ChatGPT exchanges rated high-risk, and scoring with the user's accumulated context raised that share by 13 to 17 percentage points. arXiv
  • Kairos holds an agent's untrusted side effects at the system boundary while the agent keeps running speculatively, cutting AgentDojo prompt-injection success from 47.0% to 2.7% without losing benign utility. arXiv
  • An independent reproduction of five factuality alignment methods — PPO, GRPO, Dr. GRPO, DPO and KTO — finds none significantly beats the base model, with rewards optimising format, length or brevity instead. arXiv
  • Training prompt-injection probes with matched injection-free negatives drops false alarms from 41–80% to 4–27% while still catching a median 95% of the attacks that fooled the agent. arXiv

Algorithmic advances

  • Curvature-aware pruning (CACTUS) closes 45% of the C4 perplexity gap at 50% sparsity on Qwen3-4B and runs over 28 times faster than LLM-Surgeon, while keeping the pruned model's gradients usable for training-data selection. arXiv
  • FastEEP rescales expert-parallel mixture-of-experts serving inside a running instance 4.5 to 11.7 times faster than the existing vLLM-based approach. arXiv
  • SP²ec adapts the draft configuration of self-speculative decoding during generation with a training-free bandit, adding up to 11% throughput for vision-language models. arXiv
  • A reproduction study finds self-distillation gains (SDPO, OPSD) do not generalise to unseen problems across Qwen3 1.7B–14B and OLMo-3-7B, attributing them to overfitting rather than privileged information. arXiv