The top things worth knowing about in AI today.
OpenAI has opened a limited preview of Ultrafast, a service tier running GPT-5.6 Sol at up to 750 output tokens per second, roughly 14 times standard speed, on Cerebras hardware rather than GPUs. The model intelligence is unchanged; only throughput differs. Real-time response has until now meant dropping to a smaller model, so this matters most for voice agents, support desks and anything a person sits waiting on.
Read moreDeepSeek released the general availability build of V4-Pro on 13 August, pitched at agent work: tool use, code execution and multi-step tasks. It scored 53 on the Artificial Analysis Intelligence Index, level with Zhipu GLM-5.2 but four points behind GPT-5.6 Terra, and placed 12th on the Vals Index, with notable weakness in sandboxed terminal tasks and spreadsheet modelling. Security specialists rated it higher than the aggregate scores suggest, a reminder that index rankings say little about your specific workload.
Read moreAnthropic is in early talks to acquire Israeli startup Decart for about $6 billion, which would be its largest acquisition by a wide margin. Decart builds software that makes AI chips run more efficiently, lowering the cost of training and serving models; it raised $300 million in May at a $4 billion valuation. The deal is unsigned and may not close, but the reasoning is plain enough: compute capacity, not model quality, is the binding constraint right now.
Read moreMicrosoft has started rolling out a single Copilot app replacing the separate consumer and Microsoft 365 versions, beginning with mobile and web this month and reaching Windows and Mac in September. Low-engagement features including Copilot Podcasts and Copilot Labs are being cut, and a paid AutoPilot tier for background agents is being added. If your internal documentation or training assumes the old split, it needs revising before the desktop rollout arrives.
Read moreGoogle has consolidated AI research and operations under Koray Kavukcuoglu in Mountain View, with Demis Hassabis moving to chairman of Google DeepMind and Alphabet chief scientist, working from a new London office. The 2023 DeepMind and Google Brain merger left researchers split across two continents and slowed decisions. Several senior figures have since departed, including Jeff Dean, leaving Kavukcuoglu the last Gemini co-lead still at the company.
Read moreThe federal government has taken its proposed national framework for large AI data centres to National Cabinet this month, with legislation expected in early 2027. Under the plan, large facilities would be legally required to underwrite their own new power supply, pay full connection costs, reduce consumption when the grid is under strain, and meet water efficiency standards. Anyone planning Australian AI infrastructure should treat siting and energy commitments as part of the approval process, not a later detail.
Read moreSwedish legal AI company Legora is raising at a valuation above $10 billion, roughly double the $5.6 billion it reached in March. It reports more than 1,000 enterprise customers across 50-plus markets, including Linklaters, White & Case and Deloitte. Rival Harvey is separately in talks at $15.5 billion, up from $11 billion in March. Software built for one profession is now drawing the kind of valuations previously reserved for foundation model labs.
Read moreIndian engineering group L&T won a contract worth 10,000 to 15,000 crore rupees, roughly US$1.6 billion, to build an Nvidia B300 AI factory for Together AI at its Chennai campus. The 10,000-GPU deployment will be India largest single AI compute cluster, with a first phase of 250 megawatts on a site engineered for gigawatt scale. Serious training capacity is being built well outside the usual US and European hubs, which changes where data gets processed and under whose rules.
Read moreA short read every week — the few things worth your time, and nothing that isn’t.