NAll news

All news

Tuesday792 items

  1. Unite.AIProduct release
    AI score62

    Sierra restructures the Agent Studio workflow around Ghostwriter

    On November 2, 2026, Sierra announced a redesign of Agent Studio around its agent-building tool Ghostwriter, integrating troubleshooting, journey editing, and testing into a single conversation, with evidence, suggested changes, and test results kept together for teams to review before going live. Ghostwriter was launched on March 25, 2026, and initially could read SOPs, customer service call transcripts, photos of whiteboard sketches, process documents, recordings, and natural language instructions to identify key behaviors and edge cases; on September 28, 2026, it was expanded into a proactive collaborating member that resides in Slack and Teams. The new version supports maintaining context across journeys, simulations, and dashboards; each session is a chat thread containing conversation and changes, changes must be reviewed and merged, and sessions can be shared or forked; Agent Studio can also group related agents into the same project for unified testing and release, and offers line-by-line diffs, automated Agent Checks, peer review and a Reviewer role, staged releases, and rollback-capable immutable snapshots.

    Same story · 2 reports

    #Product updates#Agents#English text is being prepared.

  2. ThurrottProduct release
    AI score60

    OpenAI plans to merge ChatGPT's Chat and Work tabs

    OpenAI plans to simplify the ChatGPT app design by merging the Chat and Work tabs, and has notified users via email. After the merge, users can bring in files and plugins in the same place, allowing ChatGPT to build websites, create presentations, or perform web operations on their behalf; Codex will remain in the desktop app's mode selector, but will share the same quota pool with ChatGPT. This interface adjustment will first roll out to a small number of users, and the shared token quota for ChatGPT and Codex will take effect for Plus, Pro, and Business users in December.

    Same story · 13 reports

    #Product updates#OpenAI#English text is being prepared.

  3. 9to5Mac
    AI score30

    iOS 27.2's New Messages Feature Suggested Actions Can Be Customized

    iOS 27.2 adds a new Suggested Actions feature to the Messages app, which can directly generate action suggestions such as adding calendar events or to-do reminders below messages. This feature is powered by Apple Intelligence and requires an AI-compatible iPhone to use. Users can turn each of the five categories of suggestions on or off separately in Settings, and all are enabled by default.

    #English text is being prepared.

  4. 9to5MacProduct release
    AI score58

    OpenAI will retire ChatGPT Classic on January 30, 2027, and is testing Chat and…

    OpenAI is notifying ChatGPT Classic users that the app will be retired on January 30, 2027; chat history will be preserved through the migration, and users are being guided to download the new desktop app. OpenAI will also test a merged Chat and Work experience with a small group of users in the coming weeks, allowing conversational features and Work's advanced capabilities to be used in the same interface, while Codex remains a separate experience within the desktop app. During the transition, OpenAI plans to introduce shared usage quotas for ChatGPT and Codex, replacing the previous separate limits.

    Same story · 4 reports

    #Product updates#OpenAI#Agents

  5. The Decoder
    AI score22

    Engineer Venkatesh Rao: AI Elites May Hoard Intelligence Like Mercantilists Hoarded Gold

    Engineer Venkatesh Rao wrote in a blog post on Contraptions, comparing the current fear of AI to mercantilism: many elites view intelligence as a scarce and dangerous resource, advocating that it be managed by experts, kept at a high price, and that rivals' access to it be restricted; some AI executives also use safety concerns to seek political influence. He believes AI will make people think more rather than less, and that the political focus should be on fair access to AI, rather than protecting traditional knowledge jobs.

    #English text is being prepared.#Industry news

  6. The Decoder
    AI score55

    Goldman Sachs revokes AI coding tool access for traders and bankers, keeping it only for engineers

    Goldman Sachs has barred most employees from using advanced AI coding tools such as Claude Code and GitHub, with only engineers allowed to continue using them, based on internal guidance from October. Traders and bankers had used these tools for several months to monitor trading flows and generate content, and the tightening is reported to have caused internal dissatisfaction. Goldman Sachs spokesperson Nick Carcaterra said the reasons include various factors such as productivity and risk management, that the trial phase has ended, and that the tools will be reserved only for those best suited to long-term use.

    Same story · 3 reports

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

  7. Sierra：BlogFirst-partyProduct release
    AI score58

    Sierra Redesigns Agent Studio Around Ghostwriter

    Sierra has redesigned Agent Studio, threading Ghostwriter throughout the entire interface so users can troubleshoot issues, modify journeys, run simulations, and test fixes within the same conversation. Each work item corresponds to a conversation, and changes remain within the conversation until review and merge, with support for sharing and forking; related agents can be grouped into the same project for unified testing and publishing, and it provides line-by-line diffs, Preview Chat, Simulations, and release approval with controlled rollout.

    Ghostwriter

    Same story · 2 reports

    #Product updates#Agents#English text is being prepared.

  9. Databricks BlogFirst-party
    AI score41

    Lakebase Postgres Generally Available on Google Cloud

    Databricks announced that its fully managed serverless Postgres database, Lakebase Postgres, is generally available on Google Cloud, supporting autoscaling, scale-to-zero, second-level branching, and point-in-time recovery. It shares the same storage layer with analytical data, eliminating the need to maintain ETL pipelines, and provides unified governance through Unity Catalog; multi-availability-zone endpoints offer a 99.99% SLA, while single availability zones offer 99.9%. Lakebase Search delivers 2x the vector and full-text search throughput of pgvector at 4x lower cost.

    #Product updates#English text is being prepared.#English text is being prepared.

  11. Thomas WolfOpinion
    AI score58

    Thomas Wolf reposted the claim that GPT-6 Astra directly rewrote the weights of Qwen3.5-0.8B

    Thomas Wolf reposted Rikiya Takehi's post, which claims that GPT-6 Astra was used to directly modify the weights of Qwen3.5-0.8B to improve its math ability; by changing only 0.016% of the weights, it fixed the model's reasoning loops without gradients or training, improving math ability by up to 6x and instruction following by 3x, and calls this approach Weight-Space Programming. Thomas Wolf himself only wrote "what?" and made no further comment.

    Quoted postRikiya Takehi · @rikiyatakehi

    We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math. It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training. We call this Weight-Space Programming: frontier LLMs can now natively read/write neural network weights! 🧵 Blog: https://t.co/bOWCcFhCVq

    #English text is being prepared.#English text is being prepared.#OpenAI

  12. Rohan PaulResearch
    AI score76

    MIT study shows GPT-6 Astra can directly modify weights to improve Qwen3.5-0.8B's math ability

    MIT researchers had GPT-6 Astra directly edit the weights of Qwen3.5-0.8B in a Python notebook, raising accuracy on MATH-500 from 7.9% to 48.1%, all without training. After Astra was given the weights and a budget of 50 practice math test attempts, it found that the model repeatedly looped until hitting the length limit, localized this habit to a single internal signal written at layer 2, and subtracted that signal from the weights. This edit changed only 0.016% of the weights, and the authors call this method Weight-Space Programming.

    Quoted postRikiya Takehi · @rikiyatakehi

    We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math. It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training. We call this Weight-Space Programming: frontier LLMs can now natively read/write neural network weights! 🧵 Blog: https://t.co/bOWCcFhCVq

    Same story · 14 reports

    #Papers/Research#Reasoning#English text is being prepared.

  15. Ma Dongxi NLPOpinion
    AI score58

    Dongxi NLP Questions GPT-6 Astra's Weight Editing Method: Like an LLM's He Jiankui

    Dongxi NLP questions the Weight-Space Programming proposed by Rikiya Takehi, arguing that the method has traces of activation steering: first find the activation direction corresponding to a certain tendency of the model, such as the loop problem of being unable to stop when solving problems, and then project out the corresponding components in the 128 columns of weights that write the most into that direction. He points out that evaluations such as MMLU-Pro, GPQA, and IFEval did indeed all improve, but these benchmarks were originally all affected by loops; once the loop is fixed, scores naturally improve across the board, and the real question is the side effects. The original post claims that GPT-6 Astra only changes 0.016% of Qwen3.5-0.8B's weights, improving math by up to 6x and instruction following by 3x, with no gradients or training required.

    Quoted postRikiya Takehi · @rikiyatakehi

    We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math. It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training. We call this Weight-Space Programming: frontier LLMs can now natively read/write neural network weights! 🧵 Blog: https://t.co/bOWCcFhCVq

    #English text is being prepared.#Reasoning#Evaluations / benchmarks

  17. Guillermo Rauch
    AI score39

    Rauch: Stateless sandboxes are the wrong benchmark; agents need to carry their workspace

    The Vercel developer account states that Vercel Sandboxes with Drives start 2.4x faster, with median creation time dropping from 1.7s to 0.7s and p95 dropping from 2.2s to 1.6s, with no SDK upgrade or configuration changes required. Vercel CEO Guillermo Rauch commented that stateless sandboxes are the wrong benchmark, and that agents are only useful when they carry their own workspace (identity, skills, repositories, memory, etc.); he also praised the Vercel Sandbox team for delivering on latency and throughput SLOs, while maintaining reliability and security amid clusters deployed across 19 global regions, new features shipping weekly, and week-over-week demand growth in the double digits.

    Quoted postVercel Developers · @vercel_dev

    Vercel Sandboxes with Drives start 2.4x faster. • Median creation drops from 1.7s to 0.7s • p95 creation drops from 2.2s to 1.6s • No SDK upgrade or config changes needed https://t.co/nuQLD0BZrp

    #English text is being prepared.#Agents#English text is being prepared.

  18. Swyx
    AI score25

    swyx shares MIT's Anqi Zhang discussing recursive language models

    swyx reposted with "lmaoooo" the Latent.Space podcast content: MIT's @a1zhang explains why Claude Code, Codex, and Pi are essentially the same, how RLM generalizes across different tasks with the help of code, context offloading, and recursive subagents, and what an OpenAI experiment with 10,000 agents and 130B output tokens reveals.

    Quoted postLatent.Space: AI News & Views · @latentspacepod

    Recursive Language Models: Claude Code, Agent Swarms, Big Research Bets, & the $40M AI Problem Solver https://t.co/VceYUKpcTo MIT’s @a1zhang explains why Claude Code, Codex, and Pi are basically the same, how RLMs use code, context offloading, and recursive subagents to generalize across tasks, why one expert can sometimes replace a trillion-token brute-force search, what OpenAI’s 10,000-agent, 130B-output-token experiment reveals about the future of language models, and why academia’s biggest advantage is the freedom to take weird, ambitious research bets.

    #English text is being prepared.

  19. Ethan Mollick
    AI score26

    Mollick: AI adoption rate continues to grow

    Ethan Mollick said the AI adoption rate is still continuing to grow, and recalled that last summer people once thought the adoption rate was declining. He also mentioned that Simon Willison had a good discussion of the methodological changes in a post.

    Quoted postSimon Willison · @simonw

    In updating a graphic showing U.S. company AI adoption over time according to a long-running U.S. Census Bureau survey, I found out that the Bureau's work has been sullied by that asinine "Super Intelligence" executive order https://t.co/YEu1tnWweT

    #English text is being prepared.

  20. Ethan MollickOpinion
    AI score62

    Ethan Mollick shares research on GPT-6 Astra directly rewriting model weights

    Ethan Mollick shared a study by Rikiya Takehi and said he is waiting for replication, but considers it a big deal. The study claims that having GPT-6 Astra directly modify the weights of Qwen3.5-0.8B to improve its math ability, changing only 0.016% of the weights, fixed the model's reasoning loops, improving math ability by up to 6x and instruction following by 3x, all without using gradients or training; the authors call this Weight-Space Programming.

    Quoted postRikiya Takehi · @rikiyatakehi

    We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math. It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training. We call this Weight-Space Programming: frontier LLMs can now natively read/write neural network weights! 🧵 Blog: https://t.co/bOWCcFhCVq

    #English text is being prepared.#English text is being prepared.#English text is being prepared.