Sources

Putting sign language AI into users’ hands deepmind.google

Introducing sign-language-to-text (SL2T), our breakthrough model powering new sign language features for Deaf and hard of hearing users.

DeepSeek V4 Pro 0813 (on OpenRouter) simonwillison.net

DeepSeek V4 Pro 0813 (on OpenRouter) The latest DeepSeek Pro model is now available, via API only. I had to link to OpenRouter because DeepSeek don’t have any obvious announcement page for their new model. I haven’t been able to confirm if they plan to release the open weights, but given the weights are available for both April’s deepseek-ai/DeepSeek-V4-Pro and July’s deepseek-ai/DeepSeek-V4-Flash-0731 it seems likely. Interestingly I got very different looking pelicans for the three different…

LFM2.5-VL-3B for Better and Faster Vision Capabilities for the Edge huggingface.co

Introducing OlmoEarth embeddings: Custom embedding exports from OlmoEarth Studio for downstream analysis huggingface.co

OlmoEarth Studio now exports custom embeddings from its Earth-observation foundation models, letting researchers plug satellite-derived vectors into downstream classifiers, clustering, and search pipelines. The workflow targets analysts who want OlmoEarth’s pretrained representations without running the full fine-tuning stack themselves.

alchemy-utils 0.1a0 simonwillison.net

alchemy-utils 0.1a0 mirrors the sqlite-utils insert, upsert, and introspection API on top of SQLAlchemy, working against PostgreSQL, SQLite, and DuckDB. Willison built the prototype in a single morning by prompting Codex and GPT-5.6, treating it as a research spike turned alpha release.

Quoting Florian Herrengt simonwillison.net

Herrengt argues teams are shipping AI-generated code no one understands, then asking Claude to debug bugs Claude itself introduced. The essay describes engineers watching walls of confident model output with no way to verify it, warning that codebases become unmaintainable layers no human owns.

There are no lossless transformations of natural-language text simonwillison.net

Alpert’s internal policy on AI writing insists engineers stand behind every sentence in their docs, because any rephrase by a model without your full intent loses information. Blaming an unclear line on the LLM, she writes, wastes reviewers’ time and misrepresents the author.

datasette-upload-dbs 0.5a0 simonwillison.net

The 0.5a0 release formalizes a POST endpoint for uploading or replacing SQLite databases on a live Datasette instance via bearer token. The workflow lets CI systems like GitHub Actions build fresh databases and hot-swap them into production once the build finishes.

References

FAR.AI security stress test far.ai

Attackers achieved success rates between 98% and 100% using relatively simple adversarial techniques… a public jailbreak developed for the model’s predecessor, V3.2, worked on V4 Pro without any modification.

SCMP — DeepSeek V4 Pro coverage scmp.com

V4 Pro struggles on benchmarks but shines in cybersecurity — outperforming Anthropic’s Opus 5 and Qwen 3.8 in vulnerability detection, though with ‘poor precision’ and elevated false positives.

prathamp.com — CSA/HCA technical writeup prathamp.com

CSA compresses KV entries at a 4:1 ratio using a learned Lightning Indexer running in FP4 to score top-k=1,024 compressed blocks, while HCA applies 128:1 compression for dense global attention.

kingy.ai — price fact-check kingy.ai

At roughly $0.435 per million input tokens and $0.87 per million output, V4 Pro 0813 is approximately 46–57× cheaper than Claude Fable 5 ($10 in / $50 out).

deepseekv4pro.com — practitioner review deepseekv4pro.com

In a Codex CLI head-to-head, V4 Pro 0813 took 12 minutes to ship a feature — four times longer than Grok 4.6 — and introduced a bug the faster model avoided; long agent loops also exhibit a ‘premature termination’ failure around round 42 of 50.

Hugging Face — DeepSeek-V4-Pro repo huggingface.co

V4-Flash moved to MIT-licensed open weights on 31 July 2026, but the V4-Pro repo still hosts only the April preview checkpoints — the production 0813 build remains API-only, with an open-weights release promised ‘soon.’

Unite.AI — AISLAC joint impact report coverage unite.ai

The committee explicitly rules out the use of AI in high-stakes professional contexts, such as medical consultations, legal proceedings, police interactions, and classroom instruction… SL2T does not fulfill the legal requirements for ‘reasonable accommodations’ under the Americans with Disabilities Act.

Deaf Journalism EU — Maartje De Meulder critique deafjournalism.eu

Developers frequently treat sign languages as datasets to be ‘extracted’ without the active participation or benefit of the Deaf community… models are trained on datasets produced by non-native signers, freezing a single, often artificial norm of signing.

Notebookcheck — Pixel 11 SL2T review notebookcheck.net

Performance degrades significantly during rapid fingerspelling—frequently misidentifying similar handshapes (e.g., ‘prey’ vs. ‘grey’)—and in low-light environments.

arXiv — MediaPipe Holistic limitations study arxiv.org

When a hand is occluded—such as being placed behind the back or hidden by the other hand—the model often ‘hallucinates’ landmarks. These phantom points do not correspond to the actual hand position and can align incorrectly with other visible body parts.

LiveLingo — competitive landscape analysis livelingo.io

In April 2026, Sorenson unveiled AI Sign Language Translation (AST) proofs-of-concept designed for quick interactions, such as retail checkouts, where a live interpreter is impractical.

SiliconANGLE — SL2T launch coverage siliconangle.com

DeepMind has not released open-source model weights for SL2T to the general public. The model is currently proprietary and restricted to the Pixel 11 hardware ecosystem… third-party developers cannot currently build upon the SL2T framework directly.

Liquid AI — LFM Open License v1.0 liquid.ai

Commercial use is free for entities with annual revenue below $10 million USD; above that threshold users must contact Liquid AI for an enterprise license.

AIModeling — Antidoom / FTPO writeup aimodeling.com

On Qwen3.5-4B the doom-loop rate under greedy sampling dropped from 22.9% to just 1% after FTPO training, without teaching the model any new skills — it simply removed the distributional pathology.

OpenReview — ScreenSpot-Pro paper openreview.net

Leading 7B models that score ~90% on ScreenSpot-v2 drop below 50% on ScreenSpot-Pro’s high-resolution professional software, indicating v2 has begun to saturate.

Hugging Face discussion — LFM2-2.6B-Exp-GGUF huggingface.co

While the 2.6B model is highly competitive in instruction following (77.83 on ToolSandbox), its coding and math capabilities remain significant weaknesses compared to Qwen 3.5.

Pau Labarta Bajo Substack — real-time video captioning with LFM2.5 paulabartabajo.substack.com

LFM2.5-VL processes a 5-frame video in just 34 milliseconds, and the WebGPU browser demo runs entirely client-side via ONNX Runtime Web with no backend infrastructure.

arXiv 2505.00684 — SigLIP2 / NaFlex arxiv.org

NaFlex processes images at their native aspect ratio and variable resolutions without the distortion caused by fixed-size cropping, significantly improving OCR, document understanding, and UI navigation.

Sources

References

Jack Sun, writing.