AtMem 2.3.10 is out; and our full DolphinBench result has more than doubled. AtMem 2.3.9 passed 192 of 600 tasks (32%). AtMem 2.3.10 passed 390 of 600 (65%). We doubled it. What changed? We focused on how an agent discovers and uses evidence. AtMem combined lexical and semantic retrieval while keeping related passages together with their original source and provenance and more. Where do we stand? This is substantial progress, with a clear tradeoff: median recorded task latency increased from 31.87 to 85.78 seconds. More successful investigation currently takes more time. The published Hermes/Luna/Mem0 reference scores 424/600 (70.7%). What comes next? We want to understand the remaining failures: was evidence missing, poorly ranked, packed incompletely, interpreted with the wrong temporal state, or used incorrectly during an action? Evidence discovery remains opt-in. Release: github.com/aetna000/atmem/re… DolphinBench submission PR: github.com/mem0ai/dolphinben… Results and execution evidence: huggingface.co/datasets/atme… PS: DolphinBench tests whether an agent can use a simulated user’s history to complete actions in applications. Finding a relevant passage isn’t enough—a task passes only when all its required checks pass.

Interesting but until it is not universal won’t be enough

What happens when an agent finishes a report, then the source spreadsheet changes before anyone approves it? This paper argues that agents need to track the work they still owe, alongside the information needed to do it. The authors bring existing research into one framework covering models, runtime harnesses, workspaces, tasks, outcomes, and verification. They distinguish two demands: how long responsibility remains open, and how much the agent must discover while working. They then organise agent capabilities, memory, workflows, training methods, and evaluation around those demands. The result is a research framework, with no newly tested agent or performance gain. Numbers throughout the paper come from earlier studies. They support parts of the discussion, but do not establish that this particular framework works better than existing approaches. Its most useful distinction is between trying an action, actually changing something, and checking that the change meets the request. That matters when an interrupted agent resumes: blindly retrying a message or payment could duplicate its effect. The discussion also asks whether accepted work proves useful and how much human checking it requires. The weakness is that the organising ideas remain unvalidated. There are no tested scales for the two dimensions, no evidence that reviewers would classify tasks consistently, and no experiment showing better reliability or lower costs. The recurring report example explains the concepts without testing them. For builders, this offers questions: which source version was used, what remains pending, and what needs rechecking? Treat it as design guidance, then test benefits. arxiv.org/pdf/2610.35845

Experience CI we will hear more about it

I can admit it won't work if you have a keyboard in between, there are some models of windows laptop made it usable by allowing to disattach the keyboar or fold it back away

may be you like the huggingface duck :) and tempted to build it yourself or buy it. 1) go here to buy it pollen-robotics.com/microduc… --> certainly very good deal since it comes with 2 IMU, camera and lidar 2) you can build it; try it here github.com/LuwuDynamics/xgod… --> does not come with uno controller, no camera or lidar, IMU simplified, I think only one if you go with 2, you need to get the kit about the same price of option 1 pollen, but as I said not the same lower quality, in particular depth and camera are real deal breaker. Any way, for those want to print the robot using the open source option 2 and buy the kit, kit is shop.xgorobot.com/products/x…

DeepSeek API now supports web search and fetch, charged at $0.75 per 1000 successful requests. Please check the details on the pricing page. Please note that when you use the DeepSeek API through DeepSeek Harness, Claude Code, or Codex, server-side tool calls have already hit the Search and Fetch APIs. These calls are currently free of charge; per-request billing starts at 00:00 Beijing time on November 1, 2026.

Computer use is not the best use of AI but if we manage it, it will supercharges the adoption

ANT IPO 🐜Read yourself sec.gov/Archives/edgar/data/… Anthropic will be listed on the Nasdaq (ticker: ANT). They don't know the price per share, the number of shares or when the final offering date will be. 50.1% of the shareholder voting power will still be held by their founders. (In general, a majority of their board seats will be controlled by their Long-Term Benefit Trust.) So you could potentially buy shares and be a shareholder, but you won't have much say in what the company does. It's growing fast. We'll see how well its shares are priced and how well they can grow and sustain profits.

when we deploy models to the production we will see it matters the model is cheap and CAPABLE

Terra 6.1 is here

it's me our you also think Claude voice is way behind ChatGPT

GTP Sol 6.1 being much powerful than 5.6 and cheaper is called progress

AtSys 0.3.0b3 — Give your agent a goal, get a verified outcome Delegating a long task should not mean watching a terminal, spotting every failure and reminding the agent what “finished” means. AtSys adds durable supervision to your coding agent. It keeps the goal, progress, recovery policy and acceptance evidence outside the conversation, so closing a terminal or losing context does not lose the task record. What you get A clear plan before execution. AtSys records the goal, deliverables, success criteria and constraints, asking for clarification when a material decision is missing. Progress you can follow. See completed work, current stages and blockers in the terminal or dashboard. Bounded automatic recovery. Diagnose failures, retry or repair within the permissions and limits you set, then escalate when human action is needed. Independent completion checks. A worker saying “done” is not enough; the declared validator decides acceptance. An automatic final report. Every completed, blocked or stopped task retains its outcome, evidence, remaining failures, available costs and next action. Completed items and passing items stay distinct. Use it for feature development, bug fixes, tests, builds, migrations, research scripts and benchmarks. Report generation runs locally without an extra model call; model-backed workers and recovery may incur provider costs. Get started python -m pip install --upgrade atsys==0.3.0b3 cd your-project atsys init Then ask Codex: $atsys Implement this feature, preserve unrelated changes, and verify it with the project’s tests. Beta: macOS and Linux; Windows supervision is unsupported. Recovery depends on your configured permissions and available tools. Release and upgrade notes · Source code

X algorithm is so interesting in particular on "For you" tab. Please use "Not interested", the impact is immediate

Whoever defines “good” will lead the development and will stay in modern software development

A100 is my go to available and value GPU for instance