Thomson Reuters ships its own in-house frontier model
Thomson Reuters has launched Thomson, a proprietary large language model trained in-house on Westlaw, Practical Law, Checkpoint, and Reuters content, after a reported $40 million investment. It's live today inside CoCounsel Legal's Tabular Analysis feature for legal and tax document review, and the company says it reached frontier-level performance on professional tasks using less than 10% of its available content. The move gives Thomson Reuters full ownership of its model, training data, and deployment — a bet that a smaller model specialized on proprietary domain content can beat general-purpose frontier models on citation-heavy legal research without third-party dependency.
Sources & depth
GPT-5.6 family lands in AWS's Kiro coding agent
OpenAI's GPT-5.6 family — Sol, Terra, and Luna — is now available inside Kiro, AWS's spec-driven coding agent, after a joint tuning effort between OpenAI and AWS. On Terminal-Bench 2.1, GPT-5.6 Terra completed tasks in Kiro at roughly 82% lower cost than before, with Kiro's structured requirements-and-design workflow feeding the model clearer context before it writes code. It's the latest leg of a months-long access-expansion push for the GPT-5.6 family — after aggressive price cuts and Replit's token-free Free Mode — into more of the tools developers already use daily.
Sources & depth
NVIDIA ships Vera Rubin NVL72 and Groq 3 LPX, claims 30x tokens-per-watt for AI agents
NVIDIA put its Groq 3 LPX inference accelerator into full production and shipped Vera Rubin NVL72, the next generation of its rack-scale AI platform, alongside Vera — the first CPU it's built specifically for AI agents. NVIDIA's own agentic-coding benchmarks (via SemiAnalysis) claim up to 30x more throughput per megawatt and 35x lower cost per million tokens versus the prior GB300 NVL72, driven by disaggregated serving, distributed KV-caching, and a faster NVLink interconnect. Early adopters include Nebius (Groq 3 LPX) and SpaceXAI, which is deploying Vera CPUs for Grok infrastructure and its in-orbit Starmind AI satellite; independent benchmarks haven't yet tested any of these numbers.
Sources & depth
Sigmoid attention fixes the soft-to-hard KV-cache eviction mismatch
A new paper finds that models trained with sigmoid attention — rather than the standard softmax — survive hard KV-cache eviction at inference with almost no perplexity loss, while softmax-gated models degrade more under the same cuts. The researchers argue this is because differentiable soft gating during training doesn't reliably predict how a model behaves once entries are physically deleted at inference, a mismatch they call "soft-to-hard." Tested at GPT-2 scale on OpenWebText, sigmoid attention beat existing KV-eviction methods like H2O and KeyDiff — a small-scale result, but a concrete architectural lever for anyone building learned cache eviction into a serving stack.
Sources & depth
A cheap repair codec closes most of the gap between INT4 and BF16 weights
A new paper proposes Activation-Weighted Seeded Residual Coding (AWSRC), a small "repair" codec that patches the errors low-bit quantization leaves behind without storing an explicit codebook — instead it stores compact seed selectors that deterministically regenerate correction bases. On Qwen2.5-3B-Instruct, adding just 0.162 extra bits per weight to an INT4 baseline closed 88% of the perplexity gap to full-precision BF16, at under 1% storage overhead relative to the full model. It's aimed squarely at teams running quantized models in production who want closer-to-BF16 quality without BF16-level memory.
Sources & depth
Reverse-engineering shows MS Paint and Photos secretly watermark "local" AI images
A reverse-engineering write-up shows that Microsoft Paint and Photos invisibly watermark AI-generated images with a server-issued GUID — even when the image generation itself runs locally on a Copilot+ PC. The researcher decrypted the local ONNX moderation model and the Watermarker.dll to trace the flow: every prompt is sent to a Microsoft moderation endpoint, which returns a unique ID that gets embedded as an invisible pixel-level watermark and mirrored in signed C2PA content-credentials metadata. Microsoft discloses remote content filtering and C2PA provenance separately, but not that the two are linked through a per-prompt identifier baked into images users think were generated entirely on-device.
Sources & depth