The top things worth knowing about in AI today.
OpenAI published the first benchmarks for Jalapeno, the inference chip it built with Broadcom, at the Hot Chips conference. It claims 1.5 to 1.9 times more work per watt than the best commercially available Nvidia systems, and 1.7 to 3.6 times lower end-to-end latency, running at 700W against roughly 1,400W. If the numbers hold in production, the cost of serving models drops for OpenAI and the pricing pressure eventually reaches everyone buying tokens.
Read moreSkild AI released S1, a robotics model that watches a single human video and then performs the task with no fine-tuning, including 10-minute jobs it has never seen. Skild says matching that accuracy with current models would take 50 to 100 hours of data collection plus retraining. The demonstrations covered ordinary work such as making coffee and flipping pancakes, and the model recovered when things went wrong mid-task.
Read moreApple announced the M6, its first processor built on a 2-nanometre process, along with the quad-die M5 Ultra. The M6 adds a neural accelerator to each of its 12 GPU cores for close to 30 per cent more AI compute than the M5, and arrives in a Mac mini starting at $899 US, shipping 22 September. The pitch is running models and agents locally instead of paying per token in the cloud.
Read moreOkta made Agent SSO generally available, registering AI agents as identities in its directory alongside human staff, with short-lived tokens instead of shared static credentials. It is available to Okta's 20,000-plus customers, with Anthropic, Atlassian, Canva, Salesforce and OpenAI among the vendors supporting the approach. Okta says only 34 per cent of organisations currently apply the same identity controls to agents as to employees, which is the gap most security teams will be asked about next.
Read moreChatGPT Work, OpenAI's agent aimed at non-engineers, is used by 98 per cent of OpenAI's own staff but only 17 per cent of business subscribers and under 1 per cent of individual subscribers. Testing found permissions and configuration confusing, and four days of casual use burned 80 million tokens, about $65 US against a $20 subscription. The gap between internal and customer adoption is the clearest signal yet that agent tooling still fails on setup, not capability.
Read moreKeenable left stealth with a $26 million US seed round led by Accel, having built a search index of more than 100 billion documents aimed at AI agents rather than people. Its API is already in production at several AI labs for both training and live queries, and the company claims responses fast enough for a voice agent to answer mid-sentence. Search infrastructure built for machines is becoming its own layer, separate from consumer search.
Read moreAmazon is shutting Mechanical Turk on 30 September, ending the crowdsourcing platform Jeff Bezos once called artificial artificial intelligence. It stopped taking new customers in July after years of decline, as models took over the small labelling jobs it was built for and data startups such as Scale AI, Mercor and Prolific took the rest. The market for paid micro-tasks has not disappeared, it has moved to firms selling specialist human data at higher prices.
Read moreA survey of developers found 80 per cent describe their use of AI coding tools as closer to dependence than advantage, with 43 per cent continuing to code after hours when they meant to stop and 39 per cent struggling to switch off. Researchers point to the loss of natural stopping points, the waits for reviews and mental walls that used to end a session. Teams measuring only output from these tools are missing a cost that shows up later as burnout.
Read morePremiers and chief ministers meet the Prime Minister in Sydney today over proposed mandatory standards for large AI data centres, which would require operators to underwrite their own new power supply, pay full connection costs, cut power when the grid needs it and meet water efficiency rules. Several states are resisting the federal push. Legislation is expected in early 2027, and the outcome sets the cost base for anyone building AI infrastructure in Australia.
Read moreA short read every week — the few things worth your time, and nothing that isn’t.