— Today's main thread: DeepSeek pushes agent infrastructure to new lows with pricing and open source, while Google tries to let agents rewrite themselves.
本日のTL;DR
DeepSeek releases V4.1-Flash-1017, improving performance while cutting task runtime and token consumption, and opens its in-house Search/Fetch API at as low as $0.75 per thousand successful requests. Google open-sources Harness², a framework that recursively rewrites the agent harness at test time, claiming significant model performance gains without weight updates. Together AI proposes Hybrid Reasoning Agents, letting LLMs reason in latent space and act through text, aiming to balance efficiency and auditability. Anthropic makes repeated moves on safety governance: granting external evaluator METR "employee-level" access and announcing it will contribute 5% of its shares to a public benefit fund.
トップニュース
1
DeepSeek releases V4.1-Flash-1017 and opens its in-house Search/Fetch API
DeepSeek releases the new model V4.1-Flash-1017, which the company says improves across multiple benchmarks over the previous V4.1 Flash while shortening task runtime and reducing output consumption and token costs, achieving a "performance-cost Pareto improvement." It also opens its in-house Search API and Fetch API at $0.75 per 1,000 successful requests, providing real-time web retrieval and Markdown content scraping for agent and RAG applications.
Why it matters: This is the first case of a model vendor directly offering search and scraping infrastructure at extremely low prices, which could reshape cost structures for agent developers and directly challenge third-party retrieval services such as Exa, Brave and Kagi.
Commenters broadly agree the pricing is far below competitors, but some doubt its anti-scraping capability, rate limits and content completeness.
Google open-sources Harness²: a framework for recursively rewriting the agent harness
Google Research releases Harness² under Apache-2.0, which recursively improves an agent's harness at test time—including system prompts, tools, memory, skills and sub-agents, the components around the model rather than the model weights themselves. Official figures show a single refinement step lifts GLM-5 on WorkBuddy-Bench from 64.2 to 75.7 (+11.5 points) with no weight updates.
Why it matters: This gives developers a path to improve agent performance without relying on a stronger model, especially suited to teams with an established model stack.
The community is broadly interested in the "same model, self-improving periphery" approach, but its effectiveness still needs validation on more benchmarks.
Together AI proposes Hybrid Reasoning Agents: think in latent space, act in text
Together AI releases HybridAgents, an architecture that lets LLM agents reason in latent space while interacting with the external environment through text. At its core is a lightweight HybridLink component that supports seamless switching between text and latent modes and maintains lossless hybrid context across multi-turn interactions via continuous working memory.
Why it matters: Latent reasoning promises lower token consumption and more consistent reasoning while retaining tool-calling ability, offering agent architectures a new balance between efficiency and auditability.
Commenters recognize the potential of latent-space reasoning, but some question whether it differs meaningfully from recurrent Transformers and note repetition-loop problems in practice.
Anthropic gives external AI evaluator METR "employee-level" access
Anthropic has allowed the nonprofit evaluator METR to embed a team in its offices with access to hardware and internal systems to assess model risk. METR can publish evaluation results directly without Anthropic's approval.
Why it matters: This is the most substantial concession yet by a frontier AI lab on third-party safety evaluation and could become a reference model for industry safety governance.
Anthropic's proposed AI safety bill: government intervention only at 50 deaths or $1B in losses
According to Politico, Anthropic is lobbying for an AI safety bill that would let the government block a model's release only if it could cause $1 billion in losses or more than 50 casualties. Evaluations would be conducted by independent third parties registered with the Commerce Department; labs may pay evaluators but must not influence their independence.
Why it matters: The bill effectively sets an extremely high trigger threshold for AI safety regulation and gives the Commerce Department near-exclusive blocking authority, raising widespread doubts about regulatory effectiveness.
Commenters broadly find the 50-death or $1 billion threshold absurdly low, questioning that it values a human life at about $20 million; others argue the bill targets frontier labs and exempts smaller organizations, making it reasonable regulation.
LEASH proposes a training-free weak-to-strong escalation framework that uses latent uncertainty to decide when to escalate a computer-use agent from a weak to a strong model and keeps monitoring after escalation, reducing cost while maintaining high success rates.
🤖LEASH is a training-free framework that uses latent uncertainty to escalate from a weak to a strong computer-use agent only when needed and monitors the strong agent afterward, reducing cost while maintaining high success.
Voxel51 releases version 3.0, integrating search, annotation and evaluation of multimodal sensor data into one platform and adding a Voxel51 Agent with 45+ MCP tools.
NVIDIA Structured Data Models offer zero-training tabular prediction: using in-context learning, pretrained foundation models predict directly on single tables or relational databases.
TinyFish makes its search API free and analyzes the real cost of AI agents reading web pages: beyond retrieval fees, model token processing is the main expense.
The Free Pascal compiler fully supports dynamic packages for the first time, letting users install components and plugins without rebuilding the entire IDE.
A practical analysis of Java sealed classes and exhaustive pattern matching; commenters recognize the progress but some think Java is still catching up to Kotlin and Scala.
Commenters broadly recognize the progress of Java sealed classes and pattern matching, but some think Java is still catching up to Kotlin and Scala.
Four large open-weight models released within 72 hours, but the only one that runs within 24GB of VRAM is NVIDIA's, sparking discussion of the practical usability of "open weights."
Namecheap is accused of seizing all of a user's domains without warning; commenters broadly advise migrating soon, though some question the credibility of the anonymous post.
Commenters broadly consider Namecheap untrustworthy and advise migrating domains soon, though some note the post comes from an anonymous Reddit account and question its credibility.
A user successfully flashed router firmware with Qwen3.8 27B, while Claude refused due to safety guardrails, sparking discussion of open-source models' safety boundaries.
Sponsor
Star
MadsLorentzen /
ai-job-search
The job search that runs on your machine. AI job application framework built on Claude Code: evaluate postings, tailor CVs, write cover letters, prep interviews. Fork it and own it.
Star
ruvnet /
ruflo
🌊 The original agent harness. Deploy intelligent multi-player swarms, coordinate autonomous workflows, and build conversational AI systems. Features adaptive memory, self-learning intelligence, federation, vector RAG integration, and native Claude Code / Codex / Hermes and many more Integrated
Star
stablyai /
orca
Orca is the place for working with a fleet of agents (e.g. Grok Bot, Claude Code). Run any coding agent with your own subscription. Available on desktop, mobile and remote runtime.
Sponsor
Star
debpalash /
VoiceStudio
VoiceStudio is the open-source, fully-local ElevenLabs alternative — voice cloning, voice design, video dubbing, dictation, transcription & audiobook creation in 646 languages.
Looks quite interesting. Lots of numbers presented, but for a single model only. Highlights: File sizes are the same as Q-quants. KLD improves substantially over Q-quants. Performance suffers slightly Creating IQ quants is somewhat slow. Asterisks apply... Appologies to Mr. Addario if I mangled anything.
Some of y'all may be familiar with AutoTrust from Hugging Face. Many of their models were trending on the front page of HF and they claimed some very impressive results. But if you looked at the org's models and clicked into likes, you may notice something a bit off: These accounts sure do look awfully realistic, no? Well let's take a look at a few of these accounts (sorry for my very poor graphic design skills): Across all these accounts, we can see a pattern: a fake, realistic name seemingly f
It's an extremely lightweight speech recognition model. 949K parameters using a weight-shared Conformer-CTC architecture Just 1.06 MB for INT8 weights Around 436× real-time on a single CPU thread 11.43% WER on LibriSpeech test-clean, using greedy decoding with no external language model Apache 2.0 licensed Any feedback, question or criticism is welcome!
Just something I've been testing over few week since Embedding Gemma 2 released which is about 330MB at Q8. Not sure if this would count as 'resource' flair... Image shows cosine similarity changes as I refine the text query (aka text embedding vector) over time. Shows how well it work & quirks it has at same time. I'll leave tl;dr version for now, 3:30 am here. If anyone is interested I'll do few more tests and make example query image like above - still have 7 more sets totaling 22448 screensh
Doska is a kanban board web and desktop app that supports offline file storage (folders with .md files) or self-hosting a sync server. When there’s a sync server, collaboration features are also supported. The self-hosted setup includes versioned Docker containers for the web server and the sync server. It looks good on mobile (PWA), and I’ve put a lot of work into making it fast, responsive and pleasant to use. There’s no login gate or any other kind of gate: it’s fully MIT open source. GitHub:
Video world models turn actions and camera motion into long visual rollouts, making video a natural interface for interactive simulation. Such interaction requires responsive generation, yet coherent rollouts must retain visual memory of what has already been generated. Chunked video diffusion models usually keep this context in growing KV caches, so later chunks become progressively slower. To address this bottleneck, we introduce Velox, a training-free attention router for region-level context
Inline rendering with document.currentScript Lean Rada describes a neat trick that takes advantage of the document.currentScript property. You can use that to drop in a tiny inline element that replaces itself when the page loads: The year is document.currentScript.replaceWith( new Date().getFullYear()) If you define a shortcut function like this: $=(...n)=>document.currentScript.replaceWith(...n) You get a tiny template language: The year is $(new Date().getFullYear()) Tags: javascript
While generalist vision-language-action (VLA) models have enabled diverse open-world manipulation, real-world robotic assistance requires two critical capabilities they currently lack: personalization and reliable execution. First, executing an instruction like "make me a sandwich" requires grounding subtask decomposition in persistent user preferences rather than relying on a single generic solution. Second, reliable deployment demands explainable safety assurance that proactively prevents haza
Generative simulation can automate robot environment construction, but physically valid environments are not necessarily useful for training. The challenge is to identify environments that address a policy's failures and supervision that helps it overcome them. We introduce CoDA, a framework in which a simulation data engine and an action policy co-evolve through a shared heterogeneous memory of environment construction, policy failures, and learning progress. The memory compiles validated diagn
I use Portainer across my whole stack but I’ve been looking to switch. Any ideas on what people recommend I was looking into Komodo but looking to hear some opinions.