Pinned Tweet
Accepted into OpenAI's Codex for Open Source: 6 months of ChatGPT Pro with Codex.
As a contributor: 32 merged PRs. Rails 14, Ruby core 8, RubyLLM 5, OpenClaw 5. Mostly bug fixes and input hardening.
Plan: use Codex on real maintenance work every day and share what holds up.

ALT Left: OpenAI email that reads "You're in. Hi Andrii Furmanets, Thanks for maintaining open source projects" and offers 6 months of ChatGPT Pro as part of Codex for Open Source. Right: "Accepted into OpenAI's Codex for Open Source. 32 merged PRs as a contributor" with a bar chart: Rails 14, Ruby core 8, RubyLLM 5, OpenClaw 5.
1
83
GPT-6.2 Astra and Claude Fable 5.2 are roughly tied on Epoch’s capability index. Yet Astra trailed by 17 points on Epoch’s own workplace tasks.
I’d use the leaderboard to pick what to try. I’d use my own repo to decide what to keep.
epoch.ai/publications/how-go…
10
Usage on Grok Bot feels much better now. My bots run all day without me watching quotas, and that changes how much I hand off to them. Thanks to the SpaceX team for making that happen.
6
If you’re building with a team of AI agents, make sure you have a domain expert on that team.
You can have agents writing code, reviewing PRs, and running tests.
But who’s making sure they’re building the right thing?
More agents ≠ better software.
8
After all the hype around OpenAI Dots, I’m finding the experience pretty frustrating.
I can’t even get my repo running and execute the same CI checks that work fine in Grok Bot. Installing Chromium and running Playwright E2E tests turns into a challenge because of sandbox restrictions.
With other AI agents, I just describe the task and they get it done. With Dots, I feel like I have to babysit every step.
Is anyone else experiencing this, or am I missing something?
16
Got access to OpenAI Dots today. I named mine Прораб 🇺🇦
First thing I did was connect GitHub and Linear. I want to see how much context it can carry from one day to the next without me having to explain the project again.
17
A green test suite is weak evidence for a bug fix if it was green before the fix.
The review note I want: this test fails on the old commit and passes on the new one. Then I can rerun the proof without trusting the review.
17
Unauthenticated RCE in React Server Components is a reminder that the Flight protocol is a server attack surface, not just a rendering trick.
I check the lockfile for react-server-dom-*, bump Next, and redeploy previews. Upgrading the react package alone is not the fix.
26
Dropping JAX/SGLang and rewriting the RL and inference stack in C/C++ is the call most teams avoid, because it pushes the release back.
But when the foundation is wrong, every model after it pays for that.
Looking forward to trying 4.8 in the soft rollout.
nitter.cf/elonmusk/status/212008…
Replying to @cradrrat
We have dumped JAX/SGLang and rewritten the entire RL and inference stack from scratch in C/C++, which took some extra time.
Training (aka pre-training) is done. Lot of cleanup was needed on the SFT data to get that mostly right. Now, 4.8 is in RL and the evals are starting to climb. We will do a soft rollout to Bot users later this week and see how it goes.
Grok 4.8 will bear no resemblance to 4.7 (thankfully). It is not yet incredible, but, this time, the foundations are built right, so I expect it to evolve to be something special.
1
39
The most useful feature in a finance app is a boring one.
Click a total, see the rows behind it.
No chart or AI summary helps when a number looks wrong. Being able to trace it does.
16
One check before you trust a bank export:
opening + money in − money out = closing
Take both balances from the statement. If it's off, read the gap:
gap = one row: missing or doubled
gap = twice one row: sign flipped
way off: often a decimal mark misread
No tool needed.
19
If AI development slows down and frontier model releases pause, it might actually be a blessing in disguise. Right now, we rarely master current models because a newer version drops every two months and changes the game. A plateau would force us to stop waiting for the next magical update and finally learn how to properly apply, embed, and squeeze real value out of the tech we already have.
21
Revolut made self-hosted GLM the default coding model for ~2,000 engineers. ~$0.30 per 1M tokens, about 1/10 of its old Anthropic/OpenAI bill. Claude Code and Codex stay for hard tasks.
The lesson isn't "open beats frontier". It's: route by task.
BREAKING: Revolut moved ~2,000 engineers onto self-hosted GLM for coding and pushed Claude Code and Codex to the exception list.
Per David Abitbol, who leads AI ops there: 56 Nvidia B200s, a blended 30 cents per million tokens, roughly a tenth of what they paid Anthropic and OpenAI. Engineering lead time improved 70%.
Here's the part that matters: the gains showed up when tokens stopped being rationed. Engineers quit saving AI for the hard tickets and pointed it at everything.
Frontier models become the exception, open weights become the default. That's a pricing problem for every lab heading to IPO.
h/t @OliverMolander
Would you move your team off frontier APIs if the bill dropped 90%?
1
41
A bank CSV isn't a format. It's a dialect.
UTF-8 or Windows-1252. Debit/credit in one column or two. "Af/Bij" in a Dutch ING export. A "Total" row pretending to be a transaction. Month names in the local language.
Parsing is the easy part. Surviving the dialects is the job.
1
39




