Manufacturing and Asset Reliability Engineer. Tech scrappy. Opinions are mine. Except the bad ones.
Oil and Gas
Joined June 2025
- Tweets1K
- Following483
- Followers87
- Likes13.3K
Can't wait for three years from now when everyone is riled up about the usage hour limits for the bi-pedal OpenAI Dot Bots being cut in half by cyborg Tibo.
I love Codex and the Codex team, but man what's going on? It was clear from the jump that the 28 days is partially a play for time, maybe more compute coming online or a model finishing training, but you get the sense something very challenging is going on behind the scenes.
It's unfortunate because Terra would appears to close an actual gap in OpenAI's lineup that was badly needed. Forcing business users to choose between Sol and Luna meant Sol would be chosen, which was overkill when a Terra would be fine.
Now, hopefully the release redistributes model usage and relaxes the compute constraints. If another part of Abilene is coming online this quarter, then one should expect OpenAI to be able to finish the year stronger.
OpenAI released a model that is intelligent, accurate, and fast. So, why is everybody upset?
Well, it's pretty simple: everyone in the AI community is currently expecting the drop of a heavily intelligent frontier model like Fable 5.5. We also know from OpenAI researchers that GPT-6 has been obsolete for months, plus OpenAI has been teasing math breakthroughs with their new frontier model. Everyone in the current AI diaspora is expecting a frontier model.
So when you release Terra, the dichotomy between what people are expecting and what they are getting is quite obvious. Even if the release itself is actually really good, innovative, and very efficient, people like me and others on X are not hungry for an efficient, intelligent, small model release. Also, the numbers are pretty close to Sonnet 5.5 in terms of cost per task.
It's pretty obvious to me why the sentiment is against OpenAI right now. It feels like we're being cheated out of frontier models, and the gap between what the labs have and what the public gets is unfortunately growing (which I know to be true because it is).
Astra - Fable
Sol - Opus
Terra - Sonnet
Luna - Haiku
I'm not sure why folks are getting worked up that OpenAI is potentially getting a Sonnet competitor into the field. That midpoint is needed to help with usage balancing, especially as a sub-agent.
I respect the opinion, but conflating the risks of AI issues today with nuclear and aviation technical systems safety overlooks the fact that those industries do so as a function of the HSE risk. Hugging Face was a big deal, no question, but it wasn't Chernobyl or AA Flight 191. If anything, it was a hollow mirror of incident of Louis Slotin and the demon core.
I believe that the labs should, and will, continue to evolve their risk management as a function of the actual risks involved, but the case being made here strikes me as an industrial risk person as out of touch with how technical safety is developed in practice.
If anything, slowing down to achieve near-perfection would hand the advantage to labs without those constraints. That's a much realer risk, with many of the consequences AI safety people fear.
New in The Atlantic:
@dgrobinson resigned this week. He was among the longest-tenured employees at OpenAI—and oversaw safety reports on 12 frontier launches.
He is very worried: “The time for trial and error is over.”
You can read his essay here:
theatlantic.com/technology/2…
I find this take to be the right balance. It's fully possible to do iterative deployment in tandem with process safety and risk management, and if there's gaps then yes the labs need to close them. This is very much a "yes, and" type situation.
David's critique is substantive and deserves careful consideration. My impression of David from our time overlapping at OpenAI is of a sober and thoughtful person who didn't approach the work from an ideological lens. This isn't someone who came in with pre-formed opinions about existential risk; he wasn't to my knowledge an EA. He was consistently productive, collaborative, and well-reasoned in his approach to safety.
The core of his critique - that professional safety engineering practices from other fields haven't taken enough root in AI safety - is substantial and correct. We're entering a phase shift where AI safety organizational practices that worked a year ago, for models of 2025 capability levels, are not sufficient to prevent serious incidents. This is a new and different world and every organization needs to uplevel accordingly.
I have much more confidence in my colleagues at OpenAI than David expresses in this article, and OpenAI does much more on this front than it gets credit for. But the actual bar for OpenAI as a whole isn't just "is it a clear field leader on alignment work, and is it successfully addressing issues as they arise?" - which is no small bar to start with - but rather, "can it earn the full trust and confidence of the public that it can safely pursue a path to superintelligence?" Because if it doesn't clear that bar, it will lose the license to operate. (Given that the public is actively debating whether to explicitly ban superintelligence and/or recursive self-improvement, I don't think this is an exaggeration.) Or, much worse, it could have a critical safety incident where the real harm is unacceptable.
That bar can only be hit by increasing the level of high-reliability safety engineering practices, embracing extreme and proactive candor around incident disclosure, and implementing third party verification that is unimpeachable from a conflicts perspective and persuasive to credible experts. (And I think it's even worth it to persuade the hostile ones.)
Also a fair take. Speaking as someone who's in an iterative deployment environment with HSE implications, I resonate with the tensions David is describing and how Miles frames it.
Ultimately, safety isn't a separate machine you build but a lubricant you have to select, monitor, and refill in the machine you're building to prevent breakdowns. My theory is that it's time for a new oil to keep with the current pace.
Didn't mean to assume bad faith, just a bit impatient when we have a few hundred thousand minutes until superintelligence, and it would have taken a fair amount of typing to respond to all the specific assumptions baked into your phrasing which I didn't want to do. My view is that companies / people in them are incentivized to make a virtue of necessity or expediency, and the term iterative deployment can be used in that way -- glorifying getting things onto the market quickly as if it's the pro safety thing to do when it's actually not. Or, at least, it is not always in those specific cases - I think there is merit in the idea that reality is the best teacher etc...which is why people have sometimes "gotten away" with the idea. It's a general purpose technology so of course there's a lot to learn. But also we already almost always know more things we could do to make each launch safer/more secure ex ante, and should more often do those things.
it makes more sense if you see dot as the beginning of a transformation of the ChatGPT/Codex experience from a user-driven, thread-based experience to primarily a dot-facilitated one
Dot can act as a router essentially, but be a frontier-level brain that ensures quality and alignment with the user. The user can still create their own threads, but that'll be for power users. Most folks will be happy to not have to deal with model and reasoning pickers to ensure results.
better UX and optimized compute per user. now if only it could be with you in the real world...
👀
My experience with dot so far has been very positive. Having it as a coordinator / right-hand agent has changed my working cadence pretty substantially.
While traveling this week, I’ve been able to use my phone to have it coordinate work through Codex on my remote PC, pull context from Teams and SharePoint, prepare scripts for colleagues, and update technical documentation with charts. It’s also helped recover an email draft and flag travel conflicts.
There’s still friction. Browser work got stuck, and the pseudo phone call is an odd UX. But being able to step away from the laptop and keep work moving has been genuinely useful.
Top DevDay highlights for me are, in order 1) dots, 2) Decisions API, and 3) 6.1 Sol. New Sol is promising but Decisions API is a legit new and potentially game changer for work use cases.
Pro tier changes sting a little, but the coming days will clarify the real trajectory.
not going to throw myself off the cliff since DevDay will shed light on more details BUT basic math says Pro user Astra usage is about to get slashed. hope I'm wrong.
Hi,
Tomorrow we are re-opening the Pro $200 subscriptions to new subscribers, but together with it we are also changing how we calculate the usage for it. In effect, if you do the math, it will net out at half the dollar in API spend compared to the old Pro $200 plan.
Now that it's said, let me explain why this is happening and why you will still get more work done than if you were on the Pro $200 subscription one month ago.
(a) We didn't want to compromise in other ways and are committing to not reintroducing the 5h limit, so that you can fully use the weekly usage when you want.
(b) On the subscription, we guarantee that over time you always get more work done and with an increasing level of quality. This means that you will continue to get more value per dollar spent as a result of models getting more efficient and us passing down the improvements in the form of API price reductions.
(c) We don't want to put an incentive on ourselves to artificially inflate the API list prices to make it look like you are getting a lot (and workaround it through discounts, etc). Instead we want to continue to both rapidly reduce prices and increase capabilities of models on the API. This week we introduced GPT-6 Sol and GPT-6 Luna at 50% of their previous price. Over time, we see prices go low enough that it makes sense for most to buy usage as needed without there being a significant gap between what you get in a subscription and what you get in the API for a dollar spent.
(d) Tomorrow, we are adding more things to the subscription that won't draw on the usage, I won't reveal what that is yet.
I wanted to be transparent before all the big announcements tomorrow. Lots of new exciting things are coming to the subscriptions that will make it super compelling, but I wanted to make sure to share this change ahead of time so you can all understand it before we shower you with good news.
Codexingly,
Tibo
Bandit retweeted
[greeting a bunch of 1800s industrial engineers who have traveled into my kitchen]
"Anyway, that's how modern packing in distillation columns works, you should really get onto refinery technology. Anyway, who wants some scones?"
"I-is that an entire roll of aluminum?"
Earlier this year I was talking to my CTO about how we needed to push hard into AI augmentation.
He reminded me that people barely even use the macro recorder in Excel.
So yeah, AI is slowly being adopted but the primary limiter has always been people's own agency.
I am so glad I didn't submit my giant radiographic piping inspection film job batch last night. 😭
This quoted post is unavailable.