Relevance 10/10Importance 10/10
Meta released two frontier models today, Muse Wonder and Muse Wonder Full, and the benchmarks are legitimately competitive. Artificial Analysis scored Wonder Full at 56 on its Intelligence Index — tied with Claude Sonnet 5.5, ahead of GPT-6 Astra and Gemini 4 Argon, just behind Claude Opus 5.5 — with standout coding numbers at 79.0 on DeepSWE 1.1 and 57.5 on SWE-Marathon. The catch: only the weaker Muse Wonder is available via API, and per-task costs on the big model roughly tripled.
Relevance 9/10Importance 10/10
Microsoft announced Azure Maia 300, a multi-chiplet inference accelerator on TSMC's N2P process with 288GB of HBM4 and 18 TB/s of bandwidth per accelerator. It promises 2.5x the memory bandwidth of Maia 200, 40% more tokens per megawatt, and up to seven times the rack-scale token throughput — 96 chips per rack, production starting 2027 at multi-gigawatt scale. Maia 300 is slated to become the primary inference system behind Microsoft's own AI products.
Relevance 9/10Importance 9/10
The Wall Street Journal reports Amazon is negotiating a roughly $7 billion acquisition of Decart, the Israeli startup whose software lets AI models run portably across different processors — plus the Lucy and Oasis world models. Anthropic reportedly walked away from a roughly $6 billion deal after due diligence. Decart last raised $300 million at a $4 billion valuation in May, so this is a sharp markup, and a deal could close within days.
Relevance 9/10Importance 7/10
CEO Cristiano Amon says unnamed AI companies have asked Qualcomm for handsets capable of continuously running models of at least 100 billion parameters by 2028. That's more than triple the ~30B MoE target of the current Snapdragon 8 Elite Extreme Gen 6, and a 4-bit quantized 100B model needs roughly 50GB of memory against today's 24GB configurations. No chip, no product date — but it tells you where the labs think inference is headed.
Relevance 8/10Importance 8/10
Bloomberg's Jason Schreier reports Epic Games pushed back its unannounced Marvel and Star Wars shooter — originally slated for March — and told the team to rebuild production around AI tooling, including Anthropic's Claude. One director's internal message reportedly said "no more hand-written code," with all art moving to full AI workflows. Epic says developers remain the creative drivers; sources say the directive did not land well internally.
Relevance 8/10Importance 7/10
Satya Nadella is arguing publicly that the industry must "assume a model is compromised and contain it from the start" — no single model for consequential decisions, tamper-proof audit logs of agent actions, third-party evaluations before release, and mandatory incident disclosure. Crucially, he wants an approved human operator able to halt any model mid-task. Microsoft president Brad Smith paired it with a line about AI being less regulated than milk and washing machines.
Relevance 9/10Importance 6/10
University of Michigan Engineering released two coordinated tools, Perseus and Kareus, that together cut training energy by as much as half with no loss of speed or accuracy. Perseus manages processor clock speeds for up to 30% savings; Kareus adds another 15 to 30% by optimizing timing and resource sharing between compute and chip-to-chip communication. The team is working with Nvidia to land it in Megatron-LM.
Relevance 9/10Importance 6/10
OpenAI's classification endpoint — built on a specialized GPT-6 Luna — is now in public beta for every developer, with image input added. It returns typed answers from a developer-defined option set in about 150 milliseconds versus 1.6 seconds for a standard Luna call, at ten cents per million input tokens and nothing for output. It's aimed squarely at content moderation, request routing, and picking an agent's next action.
Relevance 8/10Importance 6/10
The AI medical scribe company closed a $90 million Series E led by Healthier Capital, with Zoom Ventures, Philips Ventures, and Hercules Capital participating. A source pegs 2026 revenue at $60 to $80 million, with the platform now deployed across more than 450 healthcare organizations. Suki has expanded beyond note-taking into revenue cycle management and guideline lookup.
Relevance 8/10Importance 5/10
Speaking at Dell's AI Leadership Symposium, Deepgram said it's targeting voice models that run natively on customer hardware with no internet connection at all. The company already supports self-hosted and air-gapped SageMaker deployments and expanded its Flux model to ten languages this year. No launch date, supported hardware, or model size announced yet — file this under direction of travel.
