Researcher / founder @elucience, PhD on AI/LLMs for Smart Contract Security.
Melbourne, Victoria
Joined July 2012
- Tweets715
- Following3.6K
- Followers456
- Likes2.4K
It's super cool to hear how people at the labs use agents etc, but some of their strategies sound kind of impossible unless you have infinite tokens, so they don't really apply outside of the lab itself for now (at least using their models).
They need agent test suites for developer productivity and happiness that take token spend and limits into account. Like, I want to use the best model, but also not burn through 1k a day.
Maybe 1 day a week someone from the harness teams have to use reasonable external token limits with customer accessible models to understand how they get used in the wild?
I really appreciate the work OpenAI did on Sol 6.1/astra minor in this regard. Banger.
can't decide between -
a) most replies (and many posts) are written by AI
b) people have been claudecept and now write like AI
c) humans have always written such slop but I never noticed it until I spoke to intelligent sand most of the day that took on the same langauge
Peter Ince retweeted
Business Council CEO article in Aus today: "The budget capital gains, negative gearing and discretionary trust tax changes take us even further backwards, adding complexity and weakening investment incentives. ....
The data shows our productivity is at crisis levels."👇
Sol is great with goals, but I think we might need a new type of command other than queue and steer.
As steer tends to appear in the compaction and get repeated (often over many compactions) whereas it's usually just me asking a question, correcting something transient, or asking for a status update, but it gets treated as a constant reminder. Maybe a /interrupt or /now? @OpenAIDevs @jxnlco
Absolutely loving the new models Had a weird issue - sure what happened here? Codex still seems to be running, last update was a few hours ago, I think there is one waiting? @OpenAIDevs @thsottiaux
I am sure that a lot of us using /goal have some form of eval in there, it would be great to have a visible running of the current stat or last test of the eval details visible so we don't have to ask "Can I have a status update?" @OpenAIDevs Thanks!
It will be nice when more of the training sets for code contain AI native and LLM based patterns, as they still tend to over index on brittle heuristics and regex when trying to identify something, it takes a long of hand holding to stop it from just adding heuristics.
OpenAI needs some Aussie (located) staff so @thsottiaux et al don't have to get paged in the middle of the night when the servers melt lol.
I think the codex servers melted, getting this error every few turns through opencode {"type":"error","sequence_number":4,"error":{"type":"server_error","code":"server_error","message":"An error occurred while processing your request. You can retry your request, or contact us through our help center at help.openai.com ..."}.
Everyone from codex is asleep I assume, but getting frequent server overloaded errors for gpt5.4. Anyone else?
Any chance we can get Pro in codex? :) Although ideally not taking away from the normal mode usage, like a separate spark-esque usage bar. @thsottiaux @jxnlco love your work!
The only problem with the pace of change and how amazing it is to build right now, is I should be spend most of my time raising but I am having too much fun building the next thing.
Did codex change the response shape? As it's failing in the @NousResearch Hermes harness at the moment and it looks like the cause different stream expectation. Only happens using codex auth, not using gpt5.4 in open router.
You can get your agent to fix it using this prompt - OpenAI Codex auth may appear broken, but the real issue is a Responses API shape change. The Codex backend can now emit the assistant answer in stream events (`response.output_item.done` / `response.output_text.delta`) while the terminal `response.completed` object has `output=[]`. Hermes’s Codex integration wrongly assumes final `response.output` must be non-empty and rejects valid responses as “response.output is empty.” Fix by capturing streamed output items/text in `_run_codex_stream` and `_run_codex_create_stream_fallback`, then backfilling/synthesizing a minimal response before `_normalize_codex_response`. Also let `_normalize_codex_response` accept empty final output when `output_text` is present.
Would be great to have a payment system for agents that operates like a child credit card, it gets approval for a spend or sub, you approve, and then sorts out the rest of the details with the provider. Stripe is in a great position for this, they own the payment rails for many.
Peter Ince retweeted
Provenance tells the world who created something.
But it says nothing about whether the creator's judgment is reliable, whether the work holds up, or what that expertise is actually worth. Until now.
Stay tuned for an upcoming announcement.
Camp x @quadronAI