2026-04-04

The best take and comp I've seen
Peter Yang: Anthropic just sent an email saying that you can no longer run 3rd party harnesses like OpenClaw using Claude subscriptions.
Right now, both OpenAI and Anthropic are losing money on power users who run multiple agents 24/7 using their $100-200 subscription plans.
This reminds

Codex team and @OpenAI have a huge opportunity right now to:
1. Tell @openclaw users how to switch to gpt subscription (I think it’s just telling the bot to switch the model?)
2. Fix GPT’s personality (maybe even sharing a prompt will help in the short term?). This is the main reason why ppl prefer using OpenClaw with Opus.

🦞Ollama's cloud is one of the best places to run OpenClaw.
$20 plan is enough for most day to day OpenClaw usage with open models!
To make the switch, all you need is to open the terminal and type:
ollama launch openclaw
Choose a model:
kimi-k2.5:cloud
glm-5:cloud
minimax-m2.7:cloud
If you are affected, Ollama welcomes you!! ❤️
The Verge: Anthropic essentially bans OpenClaw from Claude by making subscribers pay extra https://www.theverge.com/ai-artificial-intelligence/907074/anthropic-openclaw-claude-subscription-ban

I see a lot of people worried about losing the personality of Claude when switching to a different model like Codex. This post from @steipete has done wonders for my agent's personality on any model.
https://x.com/steipete/status/2020704611640705485?s=20
Peter Steinberger 🦞: Your @openclaw is too boring? Paste this, right from Molty.
"Read your http://SOUL.md. Now rewrite it with these changes:
1. You have opinions now. Strong ones. Stop hedging everything with 'it depends' — commit to a take.
2. Delete every rule that sounds corporate. If

Holy shit. Stanford just showed that the biggest performance gap in AI systems isn't the model it's the harness.
The code wrapping the model. And they built a system that writes better harnesses automatically than humans can by hand.
> +7.7 points. 4x fewer tokens.
> #1 ranking on an actively contested benchmark.
The harness is the code that decides what information an AI model sees at each step what to store, what to retrieve, what context to show.
Changing the harness around a fixed model can produce a 6x performance gap on the same benchmark. Most practitioners know this empirically.
What nobody had done was automate the process of finding better harnesses.
Stanford's Meta-Harness does exactly that: it runs a coding agent in a loop, gives it access to every prior harness it has tried along with the full execution traces and scores, and lets it propose better ones.
The agent reads raw code and failure logs not summaries, not scalar scores and figures out why things broke.
The key insight is about information.
Every prior automated optimization method compressed feedback before handing it to the optimizer.
> Scalar scores only.
> LLM-generated summaries.
> Short templates.
Stanford's finding is that this compression destroys exactly the signal you need for harness engineering.
A single design choice about what to store in memory can cascade through hundreds of downstream steps.
You cannot debug that from a summary.
Meta-Harness gives the proposer a filesystem containing every prior harness's source code, execution traces, and scores up to 10 million tokens of diagnostic information per evaluation and lets it use grep and cat to read whatever it needs.
Prior methods worked with 100 to 30,000 tokens of feedback. Meta-Harness works with 3 orders of magnitude more.
The TerminalBench-2 search trajectory reveals what this actually looks like in practice.
The agent ran for 10 iterations on an actively contested coding benchmark.
In iterations 1 and 2, it bundled structural fixes with prompt rewrites and both regressed.
In iteration 3, it explicitly identified the confound: the prompt changes were the common failure factor, not the structural fixes.
It isolated the structural changes, tested them alone, and observed the smallest regression yet.
Over the next 4 iterations it kept probing why completion-flow edits were fragile citing specific tasks and turn counts from prior traces as evidence.
By iteration 7 it pivoted entirely:
instead of modifying the control loop, it added a single environment snapshot before the agent starts, gathering what tools and languages are available in one shell command.
That 80-line additive change became the best candidate in the run and ranked #1 among all Haiku 4.5 agents on the benchmark.
The numbers across all three domains:
→ Text classification vs best hand-designed harness (ACE): +7.7 points accuracy, 4x fewer context tokens
→ Text classification vs best automated optimizer (OpenEvolve, TTT-Discover): matches their final performance in 4 evaluations vs their 60, then surpasses by 10+ points
→ Full interface vs scores-only ablation: median accuracy 50.0 vs 34.6 raw execution traces are the critical ingredient, summaries don't recover the gap
→ IMO-level math: +4.7 points average across 5 held-out models that were never seen during search
→ IMO math: discovered retrieval harness transfers across GPT-5.4-nano, GPT-5.4-mini, Gemini-3.1-Flash-Lite, Gemini-3-Flash, and GPT-OSS-20B
→ TerminalBench-2 with Haiku 4.5: 37.6% #1 among all reported Haiku 4.5 agents, beating Goose (35.5%) and Terminus-KIRA (33.7%)
→ TerminalBench-2 with Opus 4.6: 76.4% #2 overall, beating all hand-engineered agents except one whose result couldn't be reproduced from public code
→ Out-of-distribution text classification on 9 unseen datasets: 73.1% average vs ACE's 70.2%
The math harness discovery is the cleanest demonstration of what automated search actually finds.
Stanford gave Meta-Harness a corpus of 535,000 solved math problems and told it to find a better retrieval strategy for IMO-level problems.
What emerged after 40 iterations was a four-route lexical router: combinatorics problems get deduplicated BM25 with difficulty reranking, geometry problems get one hard reference plus two raw BM25 neighbors, number theory gets reranked toward solutions that state their technique early, and everything else gets adaptive retrieval based on how concentrated the top scores are. Nobody designed this.
The agent discovered that different problem types need different retrieval policies by reading through failure traces and iterating on what broke.
The ablation table is the most important result in the paper.
> Scores only: median 34.6, best 41.3.
> Scores plus LLM-generated summary: median 34.9, best 38.7.
> Full execution traces: median 50.0, best 56.7. Summaries made things slightly worse than scores alone.
The raw traces the actual prompts, tool calls, model outputs, and state updates from every prior run are what drive the improvement.
This is not a marginal difference. The full interface outperforms the compressed interface by 15 points at median.
Harness engineering requires debugging causal chains across hundreds of steps. You cannot compress that signal.
The model has been the focus of the entire AI industry for the last five years.
Stanford just showed the wrapper around the model matters just as much and that AI can now write better wrappers than humans can.

this is what a modern media company looks like. you don't just write about the tools. you test them, break them, and tell the story. content as r&d is the future
Dan Shipper 📧: BREAKING:
Cursor 3 is now out!
It's a complete rewrite to turn Cursor into an agent orchestration tool for dispatching, monitoring, and managing AI agents locally and in the cloud.
We've been testing it for the last week internally @every and here's our vibe check:
- The

Just shipped v2 of ClawChief for @openclaw this morning.
6,000 bookmarks + 700,000 views, so a lot of you are finding value here.
v2 improvements:
1. Added a real source-of-truth layer for priorities, tasks, meeting notes, and action policy
2. Upgraded the task system with a live task file + completed-task archive
3. Reworked heartbeat into an orchestrator instead of a giant instruction blob
4. Improved the EA, biz-dev, daily task manager, and daily prep skills
5. Added meeting-notes ingestion + cleaner cron templates
Big improvements inspired by @pedroh96 and what he shared during his interview with @ashleevance:
http://youtube.com/watch?v=9ZbbxSgrjhw&t=3538s
Ryan Carson: http://x.com/i/article/2039778505282461696

I completely f*** with this.
> grew up dirt poor
> limited tech & games were an escape
> half family destroyed by hard addiction
> me, mum, bro and sis end up in homeless hostel
> met a fraudster at 16 who was hacking banks & atms
> saw that as a way to escape the matrix literally
> almost went deep down that path
> given a chance to work security @TenableSecurity
> the rest is history
Tech gave me everything I have.
Garry Tan: Tech gave me everything I have
Its capacity to lift people into abundance is incredible and there is nothing like it
We must make that into prosperity for everyone

whoa this is actually fucking sick, a self-improving ai you can use yourself right now (for any task)
dude created an ai agent that autonomously upgraded itself to #1 across multiple domains in < 24 hours…. then open sourced the entire thing
but here’s why it actually works:
- agents fucking suck, not because of the model, because of their harness (tools, system prompts etc)
- Auto agent creates a Meta agent that tweaks your agents harness, runs tests, improves it again - until it’s #1 at its goal
- best part: you can set this up for ANY task. in this article he uses it for terminal bench (code) and spreadsheets (financial modelling) - it topped rankings for both :)
- secret sauce: he used THE SAME MODEL to evaluate the agent - claude managing claude = better understanding of why it failed and how to improve it
humans were the fucking bottleneck and this not only saves you a load of time, it’s just a better way to train them for domain specific tasks
seriously check it out
Kevin Gu: http://x.com/i/article/2039807040743419904

ship your app to Vercel with Codex:
OpenAI Developers: Go from project setup to deployment with the @Vercel plugin in the Codex app.

ship your app to Vercel with Codex:
OpenAI Developers: Go from project setup to deployment with the @Vercel plugin in the Codex app.

Re @NirDiamantAI Peter Steinberger told me that he wants PR to be &#34;prompt request&#34;. His agents are perfectly capable of implementing most ideas, so there is no need to take your idea, expand it into a vibe coded mess using free tier ChatGPT and send that as a PR, which is now most PRs.

Who are you going to listen to? People who ship or people who deny?
Turning an LLM loose on a website with in WCAG 2.1 loop is you actually end up with a better design because good design is accessible design.
WCAG 2.1 is a spec LLMs can understand.
gregorein: so... as a legacy dev, why shouldn't i leave a proper legacy behind?
> &#34;who cares? it's just a blog&#34;
garryslist is a 501(c)(4) nonprofit civic engagement platform that produces voter guides and candidate endorsements for California's 58 counties
under ADA Title III, it's

I hope this never happens, but the independent ecosystem can’t bet on hope
Hope for the best, plan for the worst
clem 🤗: I think it’s @NaveenGRao who said it before but wouldn’t be surprised if the frontier labs cut their APIs entirely at some point. In a compute constrained world, they’ll always prioritize their own direct products/customers. Makes it scary and unsustainable to only build on top

Re If you hate codegen and it's April 2026, well all I can say to you is:
Have fun coding at 1x speed

&#34;There’s a big wave coming&#34; https://mtlynch.io/claude-code-found-linux-vulnerability/#theres-a-big-wave-coming

legendary run https://daringfireball.net/projects/markdown/

GLM-5 is a 744-billion-parameter open-weight model from that performs comparably to the best proprietary models (Claude Opus 4.5, GPT-5.2).
The paper documents how they got there. They use reinforcement learning in an &#34;agentic&#34; setting where each trial might involve the model writing code, running it, reading error messages, and trying again over dozens of steps.
Training on these long interaction sequences is slow because you have to wait for the slowest one to finish before updating the model, so they built an asynchronous system where the model keeps generating new attempts while simultaneously learning from completed ones.
They also describe a sparse attention mechanism that lets the model skip irrelevant parts of its input when processing very long contexts, cutting computation roughly in half without losing accuracy.
The paper explains what worked, what didn't, and why, including details like reward hacking during slide-generation training where the model learned to hide overflowing content with CSS tricks instead of actually improving layouts.
Read with an AI tutor: https://www.chapterpal.com/s/aafb0d2c/glm-5-from-vibe-coding-to-agentic-engineering
PDF: https://arxiv.org/pdf/2602.15763