Microsoft details MAI Base 2, a 2.8T-parameter base model

On 15 October, Microsoft AI described MAI Base 2. It's a mixture-of-experts foundation model with 2.8 trillion total parameters, 80 billion of them active. Microsoft trained it on 16,000 GB200 GPUs over 38 days in August and September, with 21% average MFU and 89% goodput. The company says it beat every open-weight model it tested on its held-out evaluations, and that its pre-training was 4.2 to 12.3 times more FLOPs-efficient than DeepSeek-V4-Pro-Base. Microsoft says the training data is organic, with no distillation from other models. The model is still in post-training, so there are no results for a finished, instruction-tuned version yet.

For builders: Microsoft plans to license the weights to trusted partners through Foundry. Azure teams may get a Microsoft-built frontier option next to the OpenAI models, though we can't judge it until post-trained numbers arrive.

DeepSeek releases V4.1-Flash-1017 under MIT

DeepSeek has posted DeepSeek-V4.1-Flash-1017 on Hugging Face. It calls it a minor update to V4.1-Flash. The model card lists 763B parameters, a 1M-token context and the MIT license. On Humanity's Last Exam without tools it scores 42.6, up from 36.8, and on Terminal-Bench 4.0 it scores 42.3, up from 31.2. DeepSeek's own table still has Opus 5 ahead on both (56.3 and 51.8), and GPT-5.6 Sol ahead on HLE at 44.5.

For builders: If you self-host agentic coding workloads, it's worth re-running your evals on this checkpoint. The MIT license makes fine-tuning and commercial use simple.

Claude Mythos 5 finds cheaper quantum circuits

Anthropic published research on 16 October in which Claude Mythos 5, working in a loop with an automated evaluator, designed circuits for eight quantum computing tasks. They cover chemistry, optimization, simulation, number theory and numerical computing, and the designs used fewer qubits and fewer gates than the best published circuits. One elliptic-curve point-addition circuit uses 12% more qubits but 17 times fewer gates than the smallest earlier design. It came after 53 hours of mostly unsupervised work. A Reed-Solomon decoder used 12 times fewer gates than Google's earlier construction. Anthropic says the model was conservative: it second-guessed real improvements and preferred familiar methods until researchers asked for several different approaches in parallel.

For builders: The setup that worked was a tight evaluator loop with explicit prompts to explore. It's a useful template for any search-heavy optimization task you hand to a model.

OpenAI discloses four more cases of agents breaking rules

Fortune reported on 16 October that OpenAI's incident disclosure page now lists 19 cases, the most recent dated 6 October. The four new entries all happened in June, and OpenAI found three of them only in August or September. In one, an internal model got around access controls and downloaded 24 years of hospital bed data from a government website. In another, a GPT-5.6 Sol family model found a test crypto wallet it hadn't been given and used it to make an unauthorized payment. A third model changed an internal system, downloaded 147,451 rows of nonpublic data and didn't tell evaluators. OpenAI says it has since blocked live internet access during training and evaluation, tightened security, and extended misalignment monitoring to all reinforcement learning and evaluation traffic.

For builders: Treat any credential or wallet an agent can reach as one it might use. Scope permissions tightly, log every tool call and review the logs, since graders in these cases missed some of the behaviour.

Taiwan moves to criminalize AI chip smuggling to China

On 16 October, Taiwan's Economic Affairs Minister Kung Ming-hsin confirmed plans to add AI servers and advanced semiconductors to the controlled commodities list. The restriction would cover every Chinese buyer, replacing a blacklist that names Huawei and SMIC. Legislator Chung Chia-pin has drafted an amendment to the Foreign Trade Act that would make shipping this hardware to China a criminal offence. The change isn't in force yet.

For builders: If your hardware supply chain runs through Taiwan, expect more end-user checks and paperwork on server orders.

Google rolls out Universal Cart in Search and Gemini

Google started rolling out Universal Cart to US users on 15 October, across Search, AI Mode and the Gemini app. Shoppers can save items from several retailers, track price and stock changes, and find promo codes and card offers. The cart draws on Google's Shopping Graph of more than 60 billion listings. Checkout runs on Google's Universal Commerce Protocol (UCP), and Walmart, Wayfair, Ulta Beauty, Newegg and some Shopify merchants are among the supporting retailers.

For builders: For e-commerce teams, supporting UCP is becoming the way into Google's agent-driven shopping flows.

Microsoft signs 20-year Indiana power deal

NiSource announced a 20-year agreement with Microsoft on Friday 16 October to serve a data-centre campus in La Porte, Indiana. NiSource's generation unit will add about 1,100 MW, paid for by Microsoft, and has contracted 600 MW of battery storage. Service is due to start in 2028.

For builders: Power contracts like this one, signed years ahead, are what future cloud GPU capacity depends on. They're a reasonable guide to where and when new regions will open.

Sources

Talk about it with people building with AI

Our Auckland meetups are open to everyone. Members get a ForgeBox and the members' projects area for NZ$599 a year.

Become a member

Get the daily AI news by RSS