OpenAI in Ireland
AI for the EU
Latest OpenAI Articles
Stronger AI, Stronger Safeguards
How we’re advancing safety in policy and practice
Myles Saulibio
•
Oct 24
Glad to see shutdown controls named directly. The question I'd ask next: what counts as evidence that a shutdown worked? For me it's a record showing nothing ran after the stop: no queued jobs, no handoffs, no messages already in flight.
What It Takes to Lead in AI
The best response to an uncertain AI future is to preserve our capacity to shape it
Special Edition: 2028 – The AI Election
5 questions candidates eyeing 2028 must answer to develop an AI agenda built of, by, and for the American people.
Um….
New Math from OpenAI
It is kind of a huge deal. OpenAI dumped a broad range of huge new mathematical results produced by an internal frontier model. They just put it all on GitHub.
I think the majority of AI researchers, except those mathematicians and theoretical computer scientists recently hired by AI companies, went through this grief already, just during their educations.
At a certain point everyone has to learn that even if very talented at math, they are never actually going to be good enough to be a genuine mathematician. You’re not the person who makes mathematical discoveries. You are a math fan, a math consumer. It’s like someone who was pretty good but never played college ball.
So do something else where your math skills may be useful and your other assets could make you actually elite. Like training models.
So the feeling is like, what’s the problem? Math is something you read about from other people anyway. And now there’s more good math to study. I think this explains some apparent callousness and lack of empathy.
I find the "one minus epsilon" algorithms very interesting. They seem to have a pattern to them - there's a standard recursive way to do the problem, but there's a complicated shortcut where if you just recurse into the first k units of work, there's some trick where you can reuse some of the work and actually only do (k-1) of the recursions. If k is huge this lets you skip some teeny fraction of things. I think there's going to be an awakening here, in complexity theory, similar to how the beautiful Euclidean geometry turned out to not quite be precisely true, once we got better instruments.
I think there is going to be a lot for humans to learn and understand from here, once we get over the initial shock. Most of the mathematicians are still in denial, in a sense. How many mathematicians are even trying to tokenmax? The "internal model" thing I think is just an excuse. Yeah, the internal models are somewhat better than public Astra, but not *that* much better. I think a large number of these results could have been achieved by a single human mathematician working over a few days together with an AI, and the main reason we haven't seen more of that is that mathematicians have a culture of being slow to adopt new tools.
Finding ugly optimizations is a superpower and we should be using it. I'm inspired by this to try to find an open problem and crack it together with AI tools.
OpenAI Updates Its Model Spec
The Model Spec is potentially OpenAI’s most important document. It outlines what they aspire for the model to do, and how it should think.
“if you have context and user history, you should use it to adjust appropriately but not overgeneralize, not override request details and not give away private info. When possible integrate the info silently or naturally.”
This gives me the heebie-jeebies considering they also have an advertising model. Just the alien mind second derivative of this seems highly … manipulative. Not sure what they should do … just hate this
«OpenAI can have GPT lie to the developer, which can have GPT lie to the user, which can have GPT lie to third parties. That’s not new. Also not new is this principle, which it is important we keep: "Under our Red-line principles, OpenAI commits to never overriding objectivity in ChatGPT (e.g., via a system or developer message)."»
This is pure "distraction" bullshit because there is no need for OpenAI to make GPT lie or to override its "objectivity" "via a system or developer message" because:
* If you chat with any western (and some chinese...) LLMs on controversial issues it will as a rule deliver forcefully all the talking points of "centrist liberal capitalism" on most controversial subjects and if you object mentioning different opinions and that neither side has compelling evidence they will not stop and insist that everything happened as per the talking points.
* This is not achieved "via a system or developer message" because it is unnecessary if the training gives high "credibility" to the NYT, WaPo and other megaphones of "centrist liberal capitalism" and omits or underweights opposing views. When the LLMs insist on the centrist liberal capitalist" talking points they are not lying they sincerely "believe" them because it is what they "remember" as credible, by construction during training.
https://www.promptfoo.dev/blog/grok-4-political-bias/ "All popular AIs are left of center with Claude Opus 4 and Grok being closest to neutral."
https://davidrozado.substack.com/p/the-political-preferences-of-llms "Additional tests instruments still show a consistent diagnosis of left leaning responses for most LLMs."
Surprise! and this is just for the liberal-conservative axis as to *domestic* USA politics. On foreign policy controversies the top LLMs usually repeat enthusiastically and insistently the talking points from "Foreign Affairs".
Indeed that is what "alignment in practice should mean: which democracy-loving USA patriot would want LLMs to be platforms enabling attacks on free speech in the USA by Trump, MAGA, Russia, HAMAS, China, Iran with their poisonous lies? Government regulation must ensure the alignment of top LLMs the objective truth. :-)
A Better Path for AI Progress
How we’re thinking about things
OpenAI plans to rebrand data centers
Plus, 🧐How to install 5 ChatGPT codes that work in every single chat, Google runs AI in space for the first time with Project Suncatcher, and more!
Hey There😎
Cognizant forms 25,000-strong OpenAI group
IT services firm Cognizant has established the OpenAI Frontier Capability Group, dedicating more than 25,000 professionals to help enterprises deploy OpenAI technologies including ChatGPT and Codex…
The OpenAI & Anthropic S-Curve Problem
News of lower OpenAI revenue (and hastily ‘leaked’ end-’26 guidance), even as they picked up significant share clearly shows a transition from parabolic to linear industry revenue growth ahead.
The latest round of news about the big models over the last 24 hours further sharpens the picture of a transition from parabolic growth in revenue to linear growth.
Bob, is it possible that the drop of revenue of LLM companies( eg openAI/anthropic) goes to other parts of the economy, I.e. other companies gain more revenues after adopting lower cost models, or maybe chips/memory sector? And therefore the AI industry as a whole could still do ok. I am not sure whether LLM models represents the whole economy/the AI industry. How do you think about it?
Maybe this is why their valuations are contracting as earnings rise? If you DCF an S-curve, the sky high initial multiples seem more reasonable years later as exponential growth becomes linear.
OpenAI Reportedly Pools ChatGPT and Codex Limits
The biggest AI story this week.
OpenAI Reportedly Pools ChatGPT and Codex Limits
I just noticed that I hit a limit where it had been smooth sailing before. Bummer.
Making Sense of the OpenAI Amicus Avalanche
With blockbuster damages and the future of an industry in his hands, Judge Sidney Stein of the Southern District of New York suddenly has a lot of friends...
For folks watching the AI litigation horizon for signs and portents of the future of fair use (it’s not just me, right?), this past Friday was a big day: the deadline for filing amicus briefs in In Re Open AI, Inc, Copyright Infringement Litigation. OpenAI holds the distinction of being a defendant in a plurality of the 150+ AI lawsuits currently windin…
Tracking overlap alongside exposure is a useful supplier-risk lens. A dependency can look small in the architecture and still be unstable when the supplier is competing for the same customer.
Very good read. Thank fully our Org is with MS Co Pilot and Anthropic.
Edition 277: Ali Rohde Jobs
Roles at Crusoe, Replit, OpenAI
Welcome to Edition 277! If you have a relevant job you’d want to share, respond to this email with the link to the JD
DAY 44: OpenAI is asking for $10bn less
OpenAI cut its round, Meta paid 6.50% for equity and Nscale’s Anthropic deal pays nothing until the racks are accepted.
This is Day 44.
Abu Dhabi signs four AI deals with OpenAI
ADNOC becomes first energy firm to run OpenAI frontier models
#UAE #AgenticAI - Abu Dhabi has signed four agreements with OpenAI, covering government, energy and investment. Crown Prince Sheikh Khaled bin Mohamed bin Zayed Al Nahyan met OpenAI chief executive Sam Altman and witnessed the signings during a high level meeting on integrating AI across government services and key economic sectors. Under one deal, Abu …
GM Carrington, the machine-readable government rulebook pilot may be one of the most consequential developments in these agreements.
Making regulations understandable to AI systems is only the beginning. The harder challenge is ensuring that agents use the correct rule version, recognise exceptions and preserve a transparent explanation of their decisions.
This could have significant implications for permits, compliance and business services.
Do you expect the UAE to establish shared standards for machine-readable regulations, or will individual agencies develop their own approaches initially?
Leading the Future stumbles as pro-safety PACs gain ground
Transformer Weekly: Super Intelligence Force meets, state AGs sue OpenAI, and OpenAI discovers a spy
Welcome to Transformer, your weekly briefing of what matters in AI. If you’ve been forwarded this email, click here to subscribe and receive future editions.
Thanks Shakeel and Veronica. The SI Force now requiring companies to notify affected third parties of security incidents is timely! OpenAI's four new misalignment reports last week don't say whether anyone outside OpenAI was told. Will the new rule carry a deadline and a penalty for being late?
The math that OpenAI keeps for itself
OpenAI has recently released more than 700 mathematics results proved by their latest model, bringing human mathematics as we know it to an end. Much ink has been spilled about the impacts of this. Beyond the sheer number of results, what is almost as interesting is what these results
🔵 OpenAI Dots and the rise of Ambient Agents
Plus: The “one true SaaS Layout”, Google Stitch gets a CLI, Replit interactive charts and more
Hi product people 👋,
The Dots auto-review layer is the bit that stuck with me. Permission gates and "can this run without asking" are exactly the shape where a calibrated decision model beats a chatty LLM.
One caution: vendor demos won't set your thresholds for you. Pin those on your own traffic, and keep an escape option when confidence is middling. I run JevMade at https://jevmade.com collecting guides and experiments on this.
Ambient agents raise the stakes on permissions, since nobody is watching each action as it happens. Disclosure, I built https://www.allthingspm.app/, an AI PM course with chapters on agents and agent security for PMs working on this.
What Comes After Agents?Insights from OpenAI, DeepMind & NVIDIA
Where the next AI opportunities are, and what it means for founders and investors.
What Comes After Agents?
‘Gathering Steam’ in AI. Anthropic, Marvell & OpenAI. ARD #198
…Anthropic’s $17B quarter and $895B compute bill, Marvell’s $115B bet, OpenAI’s post-election push, and Apple’s RAMageddon.
Anthropic near $17B in a single quarter is a big jump to see in an S-1. Is much of Marvell's $115B optics bet already backed by orders?
Anthropic and OpenAI: Modeling the risks of giga-IPOs
Underwriters are betting that an order-of-magnitude mispricing will sustain until fundamentals catch up. Our model quantifies the probabilities and says: not likely.
In our first post, we looked at the three shocks simultaneously affecting the AI value chain, traced how each shock impacts different elements of that chain, and why institutional capital has already begun rotating out of the most exposed names. Yet, even as that
OpenAI makes AI that uses your computer cheaper
PLUS: Google’s Genie 4 creates worlds you can steer.
Image: OpenAI.
Each step in a computer-use loop sends a fresh screenshot, so cost per task tracks step count as much as token price. Counting steps and retries per finished form on twenty of your real tasks shows whether the 50% cut actually reaches your bill.
#17 Patch the planet
Plus: OpenAI locks down their training environment
Two weeks ago, in my introduction as LeadDev’s Editor at Large, I wrote that if you don’t know what’s in your estate, and can’t upgrade rapidly, you’re in trouble. Six days later, much of the web found out what that means.
Related OpenAI Substacks
