The top things worth knowing about in AI today.
From 04:00 UTC today every deepseek-v4-pro API call is answered by V4.1-Flash, the open-weights model DeepSeek released on 10 September under an MIT licence. It runs a 552-billion-parameter mixture-of-experts backbone with a 1-million-token context window and scores 74.2 on the DeepSWE v1.1 coding benchmark, narrowly ahead of Claude Opus 5 at 74.0 and GPT-5.6 Sol at 73.0. Off-peak output tokens cost $0.60 per million, which puts real pressure on what teams expect to pay for frontier-level coding work.
Read moreSalesforce released Casey, Paige, Carter, Hunter, Marshall, Piper and Fin on 11 September, each scoped to one job rather than sold as a general assistant: returns processing, HR help desk, in-chat checkout, outbound sales, supply chain tasks. Hunter is the first to run on a new long-horizon runtime that builds work plans and reuses data across sessions, aimed at deal cycles that run for weeks. The framing is the shift worth noting: agents are being sold as headcount-shaped units of work, not as chat features.
Read moreThe NASA-IBM Lunar Foundation Model, released on 10 September, was trained on roughly 2 million image tiles drawn from nine instruments across four missions and is free to download on Hugging Face. It detects ice deposits, craters and volcanic features, cutting ice-detection error by 23% against a general-purpose vision model while using less compute. A useful reminder that a domain-specific model can beat a larger general one on the task it was built for.
Read moreAccenture and Google Cloud formed the Accenture Gemini Enterprise Business Group on 8 September, staffing it with 1,000 forward-deployed engineers drawn from a pool of nearly 50,000 Google Cloud-skilled staff. The stated goal is moving clients off pilots and into organisation-wide deployments, using industry accelerators and dedicated capability centres. The consulting layer, not the model layer, is where a growing share of enterprise AI budget now goes.
Read moreOpenAI connected ChatGPT for Healthcare to Epic on 1 September, giving clinicians access to notes, lab results, medications and specialist documentation for pre-visit review and handoffs. Epic holds records for more than 325 million patients, and physicians rated 99.1% of responses safe across 4,363 evaluations covering 27 clinical use cases. The connection is read-only: nothing is written back into the record, which is the constraint that made the integration approvable at all.
Read moreMeta's Iris accelerator enters production this month, designed with Broadcom and manufactured by TSMC under the company's MTIA program. It supports a plan to roughly double data centre compute from 7 gigawatts to 14 by 2027, against AI infrastructure spending of up to $145 billion this year. Iris runs alongside Nvidia and AMD hardware rather than replacing it, but each large buyer building its own silicon weakens Nvidia's pricing position a little further.
Read moreOxford Economics told the Australian Economic Outlook Conference in Sydney on 7 September that data centre investment will climb from $20 billion this year to $60 billion by 2030, contributing $78 billion to the economy and supporting 80,000 to 100,000 jobs a year. Data centres could draw 5-12% of national electricity within a decade. Builders want to compress construction from six years to one, and grid connection does not move at that speed, which is the real constraint to watch.
Read moreICIMS reported on 10 September that the share of workers teaching themselves AI skills rose from 22% to 30% year on year, while employer-provided training stayed flat at around 17%. Of those who say they feel ready, 61% only use general tools such as ChatGPT, Copilot or Gemini; 18% have prompt engineering skills and 17% can work on models. If your organisation is relying on staff to skill up on their own, general chatbot familiarity is what you are actually getting.
Read moreA short read every week — the few things worth your time, and nothing that isn’t.