Found this today. Pretty cool that OpenResearch's agents now read their own training curves before deciding what to try next. The search runs in rounds. The agent tries a few options for one choice, say the learning rate, and the winner becomes the parent for the next round. Each option only runs once for now, so a lucky seed can win a round (they've said as much themselves). Then the whole next round gets compared against that lucky number, so a real improvement smaller than the luck looks like a regression. And the agent's told to stop after about three regressions in a row, so one lucky winner could end a search that was still finding things. Rerunning the winner and its parent on a few seeds before promoting it is on their list, which should fix this too. Excited to see what it finds once that's in!

Nice paper, teaching agents to edit their own context mid-task with a few special tokens (NCE), rather than leaving it to the harness. On LongCat the tool-call version, trained on exactly the same edits, came in a point below plain compaction (47.3 vs 48.3), with NCE at 53.9. Though the tool version also leaves every call and its result sitting in the history, while the native edits vanish once they run. Wonder how much of the gap is just that

Great read from Sat. I think the calibration ladder is really a continual learning problem though. Brain signals drift from one day to the next, so a decoder that’s pretrained on thousands of brains and then calibrated to you once will slowly go stale. It needs to keep learning your signal as you use it, without forgetting what it picked up from everyone else. Defo worth a read!

Sonnet 5.5 improves on Sonnet 5 across benchmarks, in some cases dramatically. It’s a faster, lower-cost complement to Claude Opus 5.5, strongest at well-scoped everyday tasks, fixing bugs, and creating polished documents, slides, and spreadsheets.

the surprise conclusion of the machine consciousness debate will be that people are machines.

failing to anthropomorphize artificial intelligence is a pretty serious error of understanding that will lead you to make bad predictions and be consistently surprised by their behavior. not using terms like "thinking", "understanding", or "feeling" to describe them is a common tell that the speaker lacks deep technical knowledge of or hands on experience with modern models. if you have a friend who you notice not using this kind of language you may want to check in on them, they could have fallen into "tool-psychosis", a form of delusion where otherwise well educated people start to conceptualize sophisticated digital minds as analogous to simple computer software and find it very difficult to regain contact with reality

My rough prediction here is there's going to be two types of companies: 1. Companies that are ok with basically being an API - they expose an MCP, skills, embrace your agents 2. Companies that charge a ton for API access, try to get you using their own agents, Companies will gravitate to #2 because being API only is scary / feels very replaceable opening opportunities for #1 companies to replace them because they're more agent native I don't have a good answer for what the moat is for #1 companies but it's how I want to be using software

Will just be “meat” before too long

Is this how everyone else juggles between different subs? I'm always telling Cursor, Codex and Claude Code to pick up the conversation based on the local conversation state from another platform when I hit the limit. Works well, but it's a bit tedious.

I can't. Several years in, and no amount of prompting can prevent these models from effectively saying: "Let's add a new test suite to verify the code block I just deleted does not exist". I have several rules to discourage these endless tests, but these models are all *so* intent on LOC maxing it's unreal 😅

Dear fellow scholars, what a time to be alive

Which will happen sooner: the internet will contort to AI, or AI will contort to the internet? Both are happening. Screen understanding continues to improve massively, and over time more and more models will be trained to input and output concurrently (like @thinkymachines interaction models), enabling true continuous mouse/keyboard actions to be streamed out at the same time as the screen is streaming in (without huge multi-second blocking reasoning time preventing any true in-out flow state). On the contrary, many SaaS providers (ie @salesforce and @HubSpot etc.) have obviously focused on ensuring they can still operate in an API-first world, and are somewhat deprioritizing their UI product. My personal bet is that the internet (aka application and website authors) will continue to contort to AI. In such a world, I expect the APIs that power these apps and websites to be be effectively re-designed from the ground up, with simple stateful API-based traversel of the app/site made a priority. It's more easier to traverse a graph where the graph nodes are connected very thoughtfully with adjacent nodes having strong semantic relationships, like the connected pages of a wiki page, or the connected clickable actions on an app. REST API desgins are not really like this right now. They are stateless monolithic functions, which means models need to grep endlessly to find the right action. Navigating a website is much simpler. Think of the clickable buttons as a carefully curated set of API commands that are appropriate for your current *state* (the page you are on, or the app state you're in). SaaS providers which create fully stateful API-native versions of their app will win out in the end, and maybe the entire internet becomes a game of text-based graph traversal. This is kind of how wiki-based "brains" work today for agents, but currently this is limited to a storage system. We also need actuation of the world to follow a similar pattern, such that we keep things as intuitive for the agents as websites are intuitive for us. Just my 2 cents. Either way, it will look very different a decade from. Exciting (and strange) times ahead 👀

Great post! Quite similar to one I wrote recently: unify.ai/blog/models-are-fee… Underlines how I see agents. Also very similar to this video from @karpathy a year ago which partly inspired our work on the unify harness: piped.video/shorts/5JGcW14QY… (ignore the music, short form easier to share)

Could you cool it down with the vibe coding plz @cursor_ai team? 🙏

Very good read. Echos my feelings pretty well. The fact that these models lack true in-context learning is the biggest bottleneck imo. They’re great when starting a project (there is no context or mental model to contend with), but degrade rapidly when constructive action requires endless file grepping to *kind of* get it.