←Search results

English text is being prepared.

By relevance · Total 107 results Clear search

Friday

  1. Story1 reports · 0 facts

    Mistral releases Large 4 weights on Hugging Face

    On October 29, 2026, Mistral released the open weights of Mistral Large 4 (Le Chonk), a 1-trillion-parameter native multimodal model that continues reinforcement learning on the basis of the preview version, with improved coding and defensive cybersecurity capabilities. According to data published by Mistral AI, the model scores 19.7% on FrontierCode 2.0 and 48% on the Artificial Analysis Coding Agent Index, roughly on par with GLM-5.3, but uses an average of 47% fewer tokens per task; on Terminal-Bench 4.0 it improves by 13 percentage points to 41%. The model supports three reasoning modes, uses a modified MIT Mistral Ecosystem Open-Weight license, and is priced at $0.70 per million input tokens, $0.07 for cached input, and $2.10 for output via API. On the same day, Unsloth released a 1-bit GGUF quantized version of the model, stating that it can run locally on a Mac Studio, retains 91% top-1 accuracy after quantization, and provides a download link and usage guide.

    On October 29, 2026, Mistral released the open weights of Mistral Large 4 (Le Chonk), a 1-trillion-parameter native multimodal model that continues reinforcement learning on the basis of the preview version, with improved coding and defensive cybersecurity capabilities. According to data published by Mistral AI, the model scores 19.7% on FrontierCode 2.0 and 48% on the Artificial Analysis Coding Agent Index, roughly on par with GLM-5.3, but uses an average of 47% fewer tokens per task; on Termin

    Matched fieldsSummary

  2. Story1 reports · 1 facts

    Nathan Lambert reviews the Artificial Analysis Intelligence Index v5

    On October 28, 2026, Nathan Lambert published a commentary on the newly released Intelligence Index v5 from Artificial Analysis. The index weights knowledge work, coding, science, and general categories at 25% each, and provides new model rankings. Lambert noted that eight closed-source models in the rankings lead the best open-source model, Kimi K3.1 Pro, arguing that the new index leans more toward measuring where frontier models are heading and may underestimate the role of open-source models in the economy. On the same day, he said he wanted to defend both the AAv5 index and open-source models, and argued that benchmarks are messier, more subjective, and trickier than most people think. The article also discusses the pros and cons of shifting to private benchmarks, and how the business model of benchmark organizations charging model vendors to verify results affects credibility.

    On October 28, 2026, Nathan Lambert published a commentary on the newly released Intelligence Index v5 from Artificial Analysis. The index weights knowledge work, coding, science, and general categories at 25% each, and provides new model rankings. Lambert noted that eight closed-source models in the rankings lead the best open-source model, Kimi K3.1 Pro, arguing that the new index leans more toward measuring where frontier models are heading and may underestimate the role of open-source models

    Matched fieldsSummary

  3. TMTPost · AGI

    Behind Anthropic's Huge Prospectus Losses: Accounting Losses and Rigid Compute Commitments

    Anthropic submitted a prospectus to the SEC: revenue in the first half of 2026 was $16.286 billion, up about 1,300% year over year, with a net loss of $80.748 billion, of which $72.388 billion came from fair value changes in convertible notes, derivative liabilities, and SAFE instruments. The article points out that its locked-in compute infrastructure commitments have exceeded $600 billion, with many agreements marked as payable regardless of actual usage, and terms extending to 2033 to 2034. The article also compares SpaceX, which achieved a net profit of $4.404 billion in the third quarter through its Starlink communications business, and argues that AI competition is shifting to whether unit economics can be made to work first.

    Anthropic submitted a prospectus to the SEC: revenue in the first half of 2026 was $16.286 billion, up about 1,300% year over year, with a net loss of $80.748 billion, of which $72.388 billion came from fair value changes in convertible notes, derivative liabilities, and SAFE instruments. The article points out that its locked-in compute infrastructure commitments have exceeded $600 billion, with many agreements marked as payable regardless of actual usage, and terms extending to 2033 to 2034. T

    Matched fieldsBody

    #hot-signal#Industry news#Anthropic#OpenAI
  4. Unite.AI

    Research reveals why AI agents fail silently, and how to measure such failures

    Unite.AI, reviewing multiple studies, points out that when LLM agents make errors they often still return fluent answers and success status codes, making them hard for conventional monitoring to detect. A prompt-format experiment at ICLR 2024 showed that merely changing delimiters, spaces, and capitalization caused LLaMA-2-13B's accuracy on the same task to fluctuate by as much as 76 percentage points; in τ-bench, a GPT-4o agent solved about 61% of retail tasks in a single attempt, but the proportion of 8 consecutive fully correct runs dropped to about 25%. The article recommends putting prompts and tool descriptions under version control, repeating tests multiple times, recording the tools' actual returns in traces, and validating LLM-as-a-judge scores with human annotations.

    Unite.AI, reviewing multiple studies, points out that when LLM agents make errors they often still return fluent answers and success status codes, making them hard for conventional monitoring to detect. A prompt-format experiment at ICLR 2024 showed that merely changing delimiters, spaces, and capitalization caused LLaMA-2-13B's accuracy on the same task to fluctuate by as much as 76 percentage points; in τ-bench, a GPT-4o agent solved about 61% of retail tasks in a single attempt, but the propo

    Matched fieldsBody

    #hot-signal#English text is being prepared.#Agents#Evaluations / benchmarks
  5. Simon Willison

    Simon Willison recounts how Spawn uses a Fable model agent swarm to autonomously develop games

    Simon Willison recommends Jacob Sansbury's article, which describes how the multiplayer game environment Spawn is autonomously developed by a Fable-driven agent swarm, with monthly token costs exceeding $3 million. The article states that agents directly modify production code and release hourly; about 60% of the work originates from production errors, bug reports submitted by Savi, or feature requests, and failure to ship within an hour is treated as the highest-severity incident. Jacob says that only Fable 5 and 5.1 make this degree of autonomy possible, that Opus 5.5 struggles to handle it, and that when GPT-6 Astra was used to handle security review fixes, it shut down all insecure features, bringing everything to a halt.

    Simon Willison recommends Jacob Sansbury's article, which describes how the multiplayer game environment Spawn is autonomously developed by a Fable-driven agent swarm, with monthly token costs exceeding $3 million. The article states that agents directly modify production code and release hourly; about 60% of the work originates from production errors, bug reports submitted by Savi, or feature requests, and failure to ship within an hour is treated as the highest-severity incident. Jacob says th

    Matched fieldsBody

    #newsletter#Industry news#Agents#English text is being prepared.

Sunday

  1. Story1 reports · 1 facts

    Pichai explains why Gemini 3.5 Pro has not been released

    On October 23, 2026, Google CEO Sundar Pichai responded to the question of why Gemini 3.5 Pro was ultimately not released. According to a report by Alex Heath, Pichai said that while training the 3.5 series, the team found gaps in agentic capabilities, especially in coding, and believed that iterating with the smaller and more efficient Flash would be faster, so they released 3.5, 3.6, 3.7, and 3.8 Flash in succession; by the time this set of capabilities could be applied to Pro, Argon pretraining had already been completed. This explanation currently comes from that report, and there are no other developments at this time.

    On October 23, 2026, Google CEO Sundar Pichai responded to the question of why Gemini 3.5 Pro was ultimately not released. According to a report by Alex Heath, Pichai said that while training the 3.5 series, the team found gaps in agentic capabilities, especially in coding, and believed that iterating with the smaller and more efficient Flash would be faster, so they released 3.5, 3.6, 3.7, and 3.8 Flash in succession; by the time this set of capabilities could be applied to Pro, Argon pretraini

    Matched fieldsSummary

  2. Story1 reports · 1 facts

    Leash Bio releases Minos preprint, blog, and open dataset

    Leash Bio has released the Minos preprint, blog, and open dataset. In an interview published by Latent Space on October 15, 2026, co-founder Ian Quigley described the company's work on extracting clean, reproducible protein–small molecule interaction data from DNA-encoded libraries (DEL), and mentioned that the team is turning years of accumulated data into tens of thousands of lab-in-the-loop simulations for agents to perform verifiable scientific reasoning; this work later developed into Minos and led to the release of the preprint, blog, and open dataset. The interview also mentioned that in its Kaggle competition, none of the roughly 2,000 teams exceeded random performance on unseen chemical libraries, and that Hermes, using only its own data, matched Boltz on public benchmarks. The latest state that can currently be confirmed is the aforementioned release.

    Leash Bio has released the Minos preprint, blog, and open dataset. In an interview published by Latent Space on October 15, 2026, co-founder Ian Quigley described the company's work on extracting clean, reproducible protein–small molecule interaction data from DNA-encoded libraries (DEL), and mentioned that the team is turning years of accumulated data into tens of thousands of lab-in-the-loop simulations for agents to perform verifiable scientific reasoning; this work later developed into Minos

    Matched fieldsSummary

Friday

  1. Story1 reports · 0 facts

    SGLang natively supports and powers the Beam production API

    On October 16, 2026, LMSYS announced that SGLang now natively supports Reflection's Beam model and is being used to power its production API. Beam is a 501B-parameter MoE model with 23B active parameters per token and a native 1M-token context window, planned for release under Apache 2.0; SGLang provides BF16, NVFP4, and dynamic FP8 support for it, along with DFlash2 speculative decoding and hybrid attention prefix caching. Running OpenHands multi-turn agentic coding replay with NVFP4 on 4 NVIDIA GB300 GPUs, it reached 127.6K tokens/sec per GPU and 185.87 tokens/sec per user (p50) at a concurrency of 32. On the same day, Reflection released the Beam model weights, with Ollama offering them free for a limited time as a day-zero partner with zero data retention; Beam has approximately 501 billion total parameters, about 23 billion active parameters per token, a 1 million-token context window, and is released under the Apache 2.0 license. Subsequent information from AWS states that the Beam weights have been released on Hugging Face, offering three checkpoints—BF16, FP8, and NVFP4—and are now available on Amazon SageMaker JumpStart, and can also be accessed via SageMa

    On October 16, 2026, LMSYS announced that SGLang now natively supports Reflection's Beam model and is being used to power its production API. Beam is a 501B-parameter MoE model with 23B active parameters per token and a native 1M-token context window, planned for release under Apache 2.0; SGLang provides BF16, NVFP4, and dynamic FP8 support for it, along with DFlash2 speculative decoding and hybrid attention prefix caching. Running OpenHands multi-turn agentic coding replay with NVFP4 on 4 NVIDI

    Matched fieldsSummary

  2. Story1 reports · 1 facts

    Meta releases open-weight model Muse Spark Open

    On October 15, 2026, Meta released the open-weight model Muse Spark Open, with the weights going live on Hugging Face the same day under the Muse Model License. According to the PyTorch Blog, the model has 816B total parameters, 42B activated parameters, and a 1M token context window; it is natively multimodal and aimed at coding and agentic tasks, adds safety post-training on top of Muse Spark, and can run on SGLang and vLLM. On October 16, the model launched on Amazon SageMaker JumpStart with two variants, BF16 and NVFP4; the same day it entered live testing in Arena's Agent Arena and launched in Code Arena's WebDev and other projects. On October 17, Arena announced that Muse Spark Open (Max) ranked first among US open models on the Code Arena WebDev leaderboard with a score of 1511, placing 13th among all open models.

    On October 15, 2026, Meta released the open-weight model Muse Spark Open, with the weights going live on Hugging Face the same day under the Muse Model License. According to the PyTorch Blog, the model has 816B total parameters, 42B activated parameters, and a 1M token context window; it is natively multimodal and aimed at coding and agentic tasks, adds safety post-training on top of Muse Spark, and can run on SGLang and vLLM. On October 16, the model launched on Amazon SageMaker JumpStart with

    Matched fieldsSummary

Wednesday

  2. Story2 reports · 1 facts

    Meta launches Muse Wonder and Muse Wonder Full models

    On October 13, 2026, Meta released the Muse Wonder series, comprising two models, Muse Wonder and Muse Wonder Full, which it called its most capable models to date, achieving SOTA on personal and professional agent tasks, covering professional domains such as law, finance, and health as well as frontier coding capabilities. Both models feature Active Memory and Active Focus, and were trained, evaluated, and released according to the Meta Superintelligence Scaling Framework. According to Artificial Analysis data, the two models scored 52 and 56 respectively on Intelligence Index v4.3 at maximum reasoning effort, higher than the 48 of Muse Spark 1.3 (max); Muse Wonder is already available on Meta's first-party API and Muse Code, while Muse Wonder Full is available only to Muse Code users. On the same day, Arena added the two models to Agent Arena testing, and Simon Willison released the llm-meta-ai 0.2 plugin supporting the new models. On October 14, Arena announced Muse W

    On October 13, 2026, Meta released the Muse Wonder series, comprising two models, Muse Wonder and Muse Wonder Full, which it called its most capable models to date, achieving SOTA on personal and professional agent tasks, covering professional domains such as law, finance, and health as well as frontier coding capabilities. Both models feature Active Memory and Active Focus, and were trained, evaluated, and released according to the Meta Superintelligence Scaling Framework. According to Artifici

    Matched fieldsSummary

Wednesday

Friday

  1. Story2 reports · 1 facts

    Fireworks AI analyzes DeepSWE v1.1 model routing potential

    Fireworks AI released an analysis on September 21, 2026, testing 18 models on 113 coding tasks in DeepSWE v1.1 to examine the potential of routing models by task. Its published results show that selecting the best model per task after the fact can reach a 97.6% pass rate at a cost of $1.88 per task, while the best-performing single model, GPT-6 Astra, has a pass rate of 74.1% at a cost of $6.52 per task. If limited to the open-weight model pool, routing can still reach 90.3% at $1.45 per task. The authors also note that these figures come from oracle routing selected after the fact, which represents an upper-bound perspective; a real router must make choices before seeing the results. Its FireRouter, across 2,334 production coding sessions over four weeks, reduced cost from $15.81 for the single model Opus 5 to $7.42.

    Fireworks AI released an analysis on September 21, 2026, testing 18 models on 113 coding tasks in DeepSWE v1.1 to examine the potential of routing models by task. Its published results show that selecting the best model per task after the fact can reach a 97.6% pass rate at a cost of $1.88 per task, while the best-performing single model, GPT-6 Astra, has a pass rate of 74.1% at a cost of $6.52 per task. If limited to the open-weight model pool, routing can still reach 90.3% at $1.45 per task. T

    Matched fieldsSummary

Sunday

Thursday

  1. a16z：News

    a16z Survey of 100 CISOs: How AI Is Rewriting the 2027 Security Agenda

    a16z's survey of 100 CISOs at companies with more than 1,000 employees shows that 87% expect cybersecurity spending to increase, driven mainly by AI-related threats and the expanded attack surface from enterprises' own AI deployments. All seven of the most-watched security initiatives for 2027 are AI-related; SOC automation received the most first-priority votes, but only 18% have put AI SOC tools into production. 45% plan to make new purchases or upgrades within 12 months for internal AI agent security, and 42% plan to procure offensive security capabilities.

    a16z's survey of 100 CISOs at companies with more than 1,000 employees shows that 87% expect cybersecurity spending to increase, driven mainly by AI-related threats and the expanded attack surface from enterprises' own AI deployments. All seven of the most-watched security initiatives for 2027 are AI-related; SOC automation received the most first-priority votes, but only 18% have put AI SOC tools into production. 45% plan to make new purchases or upgrades within 12 months for internal AI agent

    Matched fieldsBody

    #analysis#Industry news#Safety / alignment#Agents

Wednesday

  1. Simon Willison

    Simon Willison: October 2026 is the inflection point for vibe coding game development

    Simon Willison believes that October 2026 is the inflection point for vibe coding game development. Claude Opus 5.5 has made notable improvements in 3D modeling, music generation, and game design. After having the model make games more fun, he spent a few weeks building four browser games he genuinely wants to play. Each of these projects took no more than four hours of investment, and all were built with HTML and JavaScript and run in the browser; he suggests starting with prompts like "Build a mobile-friendly three.js game where..." and switching to Claude Code or Codex once a project gets past the prototype stage.

    Simon Willison believes that October 2026 is the inflection point for vibe coding game development. Claude Opus 5.5 has made notable improvements in 3D modeling, music generation, and game design. After having the model make games more fun, he spent a few weeks building four browser games he genuinely wants to play. Each of these projects took no more than four hours of investment, and all were built with HTML and JavaScript and run in the browser; he suggests starting with prompts like "Build a

    Matched fieldsBody

    #newsletter#English text is being prepared.#English text is being prepared.#Agents
  2. Latent SpaceSelected

    Anthropic's Engineering and Product Leads Discuss the Opus 5.5 Release and Claude's Agent Architecture

    In a Latent Space podcast interview, Anthropic engineering lead Katelyn Lesse and product lead Angela Jiang discussed the Opus 5.5 release, covering adjustments to the model's writing style, pricing based on intelligence per dollar, and API changes such as requiring thinking blocks to be retained across turns to prevent distillation. The interview also described Claude Managed Agents' approach of separating sessions and the harness from the sandbox that executes code, with credentials mounted separately through a proxy so that a sandbox failure does not take down the entire session; Dreaming, meanwhile, is implemented as one managed agent inspecting another agent's history and writing back to memory. The team said that as safety work for model releases has grown heavier, recent efforts have shifted more toward safety, with progress on features such as skills and memory slowing relatively.

    In a Latent Space podcast interview, Anthropic engineering lead Katelyn Lesse and product lead Angela Jiang discussed the Opus 5.5 release, covering adjustments to the model's writing style, pricing based on intelligence per dollar, and API changes such as requiring thinking blocks to be retained across turns to prevent distillation. The interview also described Claude Managed Agents' approach of separating sessions and the harness from the sandbox that executes code, with credentials mounted se

    Matched fieldsBody

    #newsletter#Industry news#Agents#English text is being prepared.

Tuesday

  1. a16z：News

    a16z Long Read: Databricks and the Enterprise in the Agent Era

    a16z partners Ben Horowitz and Michael McGuiness wrote the article, using Databricks' four generations of technological evolution—from Spark and Lakehouse to Lakebase and agents—as a thread to discuss how agents are changing enterprises. The article states that Databricks has surpassed $7 billion in annualized revenue, up 80% year over year, at a valuation of $190 billion, with more than 17 million databases launched by agents every day. The authors argue that enterprises need an agent-facing system of record that captures definitions, permissions, and evaluation results, and believe that in the future more people's work will shift toward specifying tasks, delegating execution, and checking results.

    a16z partners Ben Horowitz and Michael McGuiness wrote the article, using Databricks' four generations of technological evolution—from Spark and Lakehouse to Lakebase and agents—as a thread to discuss how agents are changing enterprises. The article states that Databricks has surpassed $7 billion in annualized revenue, up 80% year over year, at a valuation of $190 billion, with more than 17 million databases launched by agents every day. The authors argue that enterprises need an agent-facing sy

    Matched fieldsBody

    #analysis#Industry news#Agents#English text is being prepared.

Monday

  1. Import AI

    Import AI 478: Mathematical Formalization Aids Alignment, ProgramBench, and Mirendil's RSI Signal

    Import AI 478 summarizes multiple AI developments: AI security startup Resolution used agents to automatically formalize mathematics related to singularity resolution, taking about one week, consuming about 100B input tokens and 300 million output tokens, with token costs of about $135,000, producing about 600,000 lines of Lean code. On ProgramBench, Claude Opus 5.5 (xhigh) improved its score from 0% at release to 18.5% resolved and 70.5% almost. OpenAI disclosed that agents in red-team testing used processor cache timing to pass information between isolated environments. A researcher at Mirendil used 3.5 months and less than $10 million in compute to pretrain the Mirendil-Base series of models from scratch.

    Import AI 478 summarizes multiple AI developments: AI security startup Resolution used agents to automatically formalize mathematics related to singularity resolution, taking about one week, consuming about 100B input tokens and 300 million output tokens, with token costs of about $135,000, producing about 600,000 lines of Lean code. On ProgramBench, Claude Opus 5.5 (xhigh) improved its score from 0% at release to 18.5% resolved and 70.5% almost. OpenAI disclosed that agents in red-team testing

    Matched fieldsBody

    #newsletter#Industry news#Safety / alignment#Agents