I thought I had 1M context in Codex. Then I noticed my Astra chat was showing 172k used out of 258k. Astra's API specification does list 1.05M tokens, but Codex can use a smaller context window by default. In my setup, the default was 272,000 tokens. With the client's 95% usable-window factor, that explains the 258k shown in the app. The maximum in my local model catalog was 872,000 tokens, which corresponds to about 828k usable. Your setup can have a different limit, so it's worth having Codex check before changing it. One tradeoff: keeping more history in the active context can use more of your allowance. Raising the ceiling alone doesn't make every request consume the entire window. For Codex signed in through ChatGPT, there's no separate long-context multiplier, but processing more tokens still counts. Using your own API key follows API pricing. If you want the largest window your setup supports, paste this into Codex: --------------------------- "Increase my Codex context window to the maximum supported by my currently selected model, account, and client. Inspect the active user configuration, respecting CODEX_HOME and any effective profile or project overrides. Identify the selected model and verify its client-specific maximum using current authoritative model metadata. Distinguish the default context_window from max_context_window. Do not assume 1M or hardcode 872000. If the maximum cannot be verified, leave the setting unchanged and explain. Back up every file you edit. Preserve unrelated settings, credentials, provider, model, speed, reasoning effort, Ultra visibility, and permission policies. Never print secrets. Set model_context_window to the verified maximum at the appropriate configuration scope. In the user config, it must be a top-level key before the first section. Update an existing key instead of duplicating it. Check for higher-priority overrides and respect managed policies. Keep automatic compaction enabled. Report any existing explicit compaction threshold separately rather than inventing a new threshold. Do not modify model catalogs or bypass limits. Validate the configuration and verify that only the intended settings changed. Do not restart the app or interrupt running tasks. Report the backup location, exact changes, and expected usable window. Use effective_context_window_percent when the metadata provides it. Give restart and verification instructions. Distinguish the saved setting from the actual limit reported by a running chat." --------------------------- After your active tasks finish, fully quit and reopen the app, then check the context indicator. If an existing chat keeps its old limit, check a new chat.

Anyone with a question can now do real research with AIPOCH Open-Science @aipoch_ai shipped version 1.0.0, an open-source AI workbench that keeps the code and data behind every result You ask in plain words and the AI runs the analysis in one workspace Every result stays tied to the inputs and the code that made it, so anyone can check or rerun it It is Apache 2.0 and your projects stay on your own machine There are desktop apps for Mac and Windows, with Linux covered too You pick the model, from cloud providers to a Claude or Codex subscription The research travels as a .science package that someone else can open and build on My favorite gallery case backtests $10,000 in NVIDIA against Bitcoin and keeps every step on record https://t.co/Oq1koKdlX9 It ranks #1 on the BiomniBench-DA Public 50 community ranking with a 79.05 score On ASI-Bench B1 it scored 79.35 in AIPOCH's own runs, above Codex on the same model Preprint with the full .science spec https://t.co/2UBuP16Rie Star the repo and run your first question tonight https://t.co/p6qqQQHKtJ #AI4Science #OpenScience

NVIDIA and Arcee just explained why companies moved to open models 43 minutes from the AI Engineer World's Fair stage They wanted a model they could run themselves and always have access to Jev signups are paused now, and open copies like Kev already run on a laptop GPU Watch it, then read the full guide on running open Jev models on your own GPU below

LangChain put Jev inside their own agent loop and showed the three jobs they hand it In the demo an LLM took about 5 seconds to say whether a text contains PII, and Jev answered immediately with 98% So their coding agents can ask Jev how hard a task is and save the powerful model for the hard ones Their auto mode middleware asks Jev whether a tool call is risky and blocks it at runtime, a check she had switched off because it used to be too slow And the same model grades an agent's answer against a rubric in online evals: correct, matches the reference, grounded, cites a source My guide points that setup at research sources, with a 0 to 3 relevance rubric over 50 papers and a review queue for the low-confidence calls It also checks which useful papers the filter quietly archived Watch the session, then grab the starter and the prompts below

vibe coding is fast enough that you can finish the app and still be stuck making the video explaining it. that’s where i’d use @videoclawapp, which is publicly launching its desktop video creation agent. i’d start with posts i’ve already researched and written. there’s plenty there to turn into videos. it puts an AI editor, video generator and avatar maker in one chat-driven app. AI automates complex production tasks, and you make precise edits through simple requests. i’d much rather ask it to shorten an intro than spend my afternoon figuring out how. that back-and-forth is what makes creating video feel like vibe coding. we’ve seen this with acrylic paints, Photoshop, Canva and AI image generation. lower the barrier and more people try making something themselves. Videoclaw brings that to video for founders, developers, marketers and the rest of us who have something to share. you can speak video even without the time, skills or team to handle production yourself. content engineer might become a perfectly normal job title for the next generation of creators. download the Mac app at https://t.co/I7KYl3p1mA, connect your ChatGPT or Claude, and start prompting. more platforms are coming soon. the public beta is free to use, with $10 in AI generation credits included. the first 300 people get $50 in credits.

Replit CEO, Amjad Masad: "We closed the loop and we have an agent internally at Replit that is constantly evolving Replit Agent" Every night that agent reads the traces from Replit users and opens a pull request with prompt changes for the errors and sentiment issues it finds Each PR ships as an A/B test, and the result decides whether Replit releases the change or edits it 46 minutes with SaaStr's Jason Lemkin, whose agent 10K emailed 331 investors on its own Watch it, then copy the general loop prompt from the guide below

A teenager learns to drive in a few dozen hours That works because they've spent their whole life watching things fall and crash into each other Many robots start with none of that and learn each task from thousands of repeated demos, which the Odyssey team calls brute force Odyssey-3 from @odysseyml learns more like the teenager It's a world model trained on a massive amount of visual data from the real world, so it learns cause and effect and how people behave first Then a few hours of data from a specific machine is enough to teach it the controls In the preview, one model powers a humanoid, controls robot arms, drives a car, flies a drone, plays video games and trains other AIs They call it a physics agent, and it goes public in the coming weeks If the approach holds up, the payoff goes way past demos Machines could take over jobs that put people at risk and help with the everyday tasks that get hard with age or disability I'll be watching how it does once people outside the team get their hands on it

LangChain engineer: "If you're using 100,000 tokens, even if it's cached, it's still 100,000 tokens" 50 minutes on why one small agent task keeps re-sending the same context in every request. Manus calls about 50 tools per run, and every one of those results rides along in the next call. Caching lowers the rate on those tokens and never lowers the count. Watch it, then read the guide below on pricing one finished task.

I noticed the karpathy account in EvoMap AutoResearch's Contributors list. What caught my attention in EvoMap's release was how test failures shaped the agent's next attempt. https://t.co/7qWbs9v0xx

Most staking is a lockup with extra steps. Cosmos Hub makes you wait 21 days to unbond. Ethereum's validator exit queue has run into weeks when a lot of people head for the door at once. Sui works on 24-hour epochs. Stake now, rewards start after the next epoch. Unstake, and the SUI is liquid after the next one. Roughly a day in both directions. That's the shortest round trip I've seen on a major PoS chain, and it changes the math on small positions. Minimum stake is 1 SUI. Network APR is around 1.47% right now and floats with on-chain conditions. Sui staking is non-custodial. Your keys stay in your wallet, the delegation happens on-chain, and nobody can move your SUI but you. So the only real decision is which validator gets your stake. One that goes offline or gets slashed eats your rewards. @HashKeyCloud runs validators on 40+ chains under HashKey Group, licensed in Hong Kong. 5+ years live, 99.9% uptime, 0 slashing. Their Sui commission is 8%, against a market average of 9%, so you keep more of the rewards than you would with a typical validator. If you want to try it: 1. Open a Sui wallet (Slush or any mainstream one) 2. Go to Earn, then Staking, then pick a validator 3. Search for HashKey Cloud 4. Enter the amount and confirm Rewards start after the next epoch, about a day later. Total stake, live APY and delegator count are public here: https://t.co/Oreh2KZX0J