AI Daily · 2026-08-13

Today's releases span models and infrastructure: DeepSeek V4 Pro 0813 quietly appeared via API, Google DeepMind brought the sign-language-to-text model SL2T to Pixel, and LiquidAI launched a 3B vision-language model for the edge. On research, Anthropic's meta-analysis of 56 US randomized studies finds worker retraining programs deliver only modest gains of about 2 to 3 percentage points in employment and would likely be overwhelmed by large-scale AI-driven job displacement, so the authors argue for investing in and rigorously evaluating the most promising projects. Google Research shows frontier LLM factual errors are more often a recall failure than an encoding failure, with lost keys rather than empty shelves as the better metaphor. Tooling and agent infrastructure are also moving, with LangChain's LangSmith BYOC generally available on AWS and managed deep agents pushed as the next stage, while Simon Willison's alchemy-utils and a quoted critique highlight both new utility and cognitive debt in AI-assisted coding.

North America · First-hand

Anthropic

⭐⭐⭐ [Research] Reviewing the evidence on worker retraining programs

Anthropic's Economic Research team, together with independent researcher David Roodman and Maxim Massenkoff, released a review of the evidence on worker retraining programs, asking whether such policies could mitigate large-scale labor market disruption from AI. Drawing on a new meta-analysis of 56 randomized US studies and experimental evidence from Europe, the report finds that job training programs produce positive but modest average effects: each offered training slot raises employment by two to three percentage points and earnings by roughly $1,000 a year, against a cost of about $13,000, while the government recovers more than half of spending and programs roughly break even overall. A small set of sector programs that partner with employers and place people directly into jobs yields gains several times larger, but attempts to replicate them have often failed. The authors conclude that existing retraining programs would likely fall short if AI displaces workers at scale, and they recommend investing now in demonstrating, evaluating, and scaling the most promising programs, including rapidly expanding one leading program and rigorously measuring the results.

⭐⭐ [Product Update] 2.1.231

Claude Code released versions 2.1.231 and 2.1.229, fixing many stability and compatibility issues such as MCP OAuth redirect URI mismatches, partially disappearing long streaming responses, crashes on Windows extended-length or UNC paths, and auto mode failing when the attribution header is disabled. New features include plugin marketplace command sources, server-supplied hook support for self-hosted runner sessions, and remote-control --continue for resuming the most recent session. Other improvements cover workflow fan-out prefix caching, clearer 'prompt too long' errors, sandbox IPv6 literal handling, and telemetry export fixes.

Google

⭐⭐⭐ [Model Release] Putting sign language AI into users’ hands

Google DeepMind has introduced SL2T, a sign-language-to-text model described as a breakthrough in quality and generality. The model brings sign language AI from the lab into consumer products for the first time, powering sign-to-text dictation in Gboard and Live Transcribe on Pixel 11, starting with American Sign Language (ASL) to English; more devices and languages are planned. Users can sign to search, draft messages or documents, ask Gemini to perform tasks, or respond in conversations instead of typing. The post explains that sign language translation requires true machine translation and visual understanding of simultaneous movements of the hands, arms, torso, head, and face. SL2T is trained on over 100,000 hours of data across more than 50 sign languages.

⭐⭐⭐ [Research] Empty shelves or lost keys? Recall is the bottleneck for parametric factuality

Google Research introduces a knowledge profiling framework that separates whether LLM factual errors stem from facts never being encoded or from encoded facts being hard to recall. The study presents WikiProfile, a benchmark of 2,150 Wikipedia-derived facts, each paired with ten questions probing encoding, recall, and recognition. Analysis shows frontier LLMs such as Gemini 3 and GPT-5 encode nearly all facts but struggle to recall many of them, making the bottleneck more like lost keys than empty shelves. Facts are classified into five profiles: encoding failure, recall failure, direct recall, recall with thinking, and inference without encoding.

Ecosystem & Beyond (Products / Agents / Tools / Opinions)

Model Release

⭐⭐⭐ [Model Release] LFM2.5-VL-3B for Better and Faster Vision Capabilities for the Edge

LiquidAI released LFM2.5-VL-3B, a vision-language model designed to run on local hardware, with strengths in document and screen understanding, object grounding, multi-image input, and function calling; it answers directly rather than reasoning to keep latency low. The model pairs a SigLIP2 400M NaFlex vision encoder with a 2.6B text backbone, is pre-trained on about 34T tokens with 4x more vision data than before, and expands its vocabulary to 128K for non-Latin scripts. Training proceeds through supervised fine-tuning with knowledge distillation and Antidoom, followed by multi-reward reinforcement learning. Benchmark results show it leads its size class on many vision and text tasks and reads documents, charts, and on-screen UI elements well.

⭐⭐⭐ [Model Release] DeepSeek V4 Pro 0813 (on OpenRouter)

Simon Willison reports that DeepSeek's latest Pro model, DeepSeek V4 Pro 0813, is now available via API only, with no obvious official announcement page from DeepSeek, so he linked to OpenRouter. He has not confirmed whether open weights will be released, but considers it likely given that weights for April's V4-Pro and July's V4-Flash-0731 are public. He also noticed that the model produced very different pelican images across low, medium, and high reasoning levels, a behavior he has not seen in other models. According to his account, benchmark results were shared in DeepSeek's official WeChat group, then pasted to Reddit where moderators removed the post as low-effort, before appearing in an ASCII-art table on Hacker News.

Product Update

⭐⭐⭐ [Product Update] LangSmith BYOC on AWS is generally available

LangChain announced that LangSmith Bring Your Own Cloud (BYOC) is now generally available on AWS. Enterprise customers can run a managed LangSmith deployment in their own AWS account and VPC, keeping sensitive data such as traces, datasets, prompts, agent deployments, and sandbox data within their cloud boundary. BYOC combines the data residency and network isolation of self-hosting with the operational experience of a managed service, targeting regulated and security-sensitive sectors like financial services and healthcare. The control plane runs in LangChain's cloud and does not hold sensitive application data, while the data plane runs in the customer's AWS account, with communication over AWS PrivateLink. The offering is available to enterprise customers across 15 AWS regions in the US, EU, and APAC.

⭐⭐ [Product Update] Introducing OlmoEarth embeddings: Custom embedding exports from OlmoEarth Studio for downstream analysis

AllenAI's OlmoEarth Studio platform now lets users compute and export embeddings generated by the open-source OlmoEarth foundation models, via the UI or API. The exports are computed on demand rather than pulled from a pre-computed archive, with options for area of interest, time span of 1-12 monthly periods, encoder variant (Nano/Tiny/Base), spatial resolution (10/20/40/80 meters), and imagery sources such as Sentinel-2 L2A and Sentinel-1 RTC. Results are delivered as Cloud-Optimized GeoTIFFs, with one band per embedding dimension and vectors stored as int8, along with a dequantization method to recover floating-point values. The embeddings support downstream tasks including similarity search, segmentation, and unsupervised exploration, and supervised fine-tuning is available for higher performance. Source code and model weights are publicly available.

⭐⭐ [Product Update] alchemy-utils 0.1a0

Simon Willison released alchemy-utils 0.1a0, an alpha cross-database version of sqlite-utils built on SQLAlchemy. The tool preserves sqlite-utils' core API, including insert, upsert, insert_all, upsert_all, create, update, and table introspection, and supports PostgreSQL, SQLite, and DuckDB. He built the prototype with Codex and GPT-5.6 Sol Ultra using very few follow-up prompts. The post demonstrates querying a local PostgreSQL blog database and inserting San Francisco street tree CSV data into DuckDB, with an optimization from nearly an hour down to about 35 seconds.

Opinion

⭐⭐ [Opinion] Why managed agents are the next big thing in agent building

LangChain launched Managed Deep Agents, described as the easiest way to build, run, and deploy production agents. Harrison Chase reviews the stages of agent building, from early AI frameworks/apps to more mature frameworks like LangGraph, Google ADK, and Vercel's AI SDK, and then to tool-calling agent loops and harnesses such as Claude Code, Pi, and Deep Agents. He argues that primitives like durable execution and sandboxes, along with standards such as AGENTS.md, MCP, and skills, are enabling managed agent experiences. By bundling harnesses with infrastructure, managed services reduce the burden of runtime, UX, sandboxing, context management, evaluation, memory, and authentication for production agents.

⭐⭐ [Opinion] What is an AI agent?

LangChain defines an AI agent as a system that uses a large language model to decide the control flow of an application, with the degree of agency depending on how much control the model owns. The post frames autonomy as a spectrum, ranging from simple LLM routers to fully autonomous systems, and proposes a six-level scale inspired by autonomous driving. It distinguishes workflows, which orchestrate models and tools through predefined code paths, from agents, where the LLM dynamically directs next steps at runtime. Citing Anthropic's definition, it also notes that giving the model more control requires stronger infrastructure for observability, evals, memory, permissions, and safe execution.

⭐⭐ [Opinion] Quoting Florian Herrengt

Simon Willison quotes Florian Herrengt describing a team that repeatedly asks AI to fix the same bug without success, and a developer who does not know where the data comes from and turns to Claude for help. The two watch an endless stream of generated text, unsure whether any of it is true. The quoted piece argues that the project has become so layered and complex that no one on the team can understand it, reflecting cognitive debt and the loss of a middle tier in software engineering.