Small Businesses, Bigger Capabilities
SMB use of agentic AI is growing fast
Teri R. Moten, MBA, CAIO
•
Oct 3
Such an honor having one of my use cases with ChatGPT Work featured!
Latest OpenAI Articles
70M Prompts About Faith
How AI, as a tool for people, is helping people explore their faith
What is really exciting is AI putting the final nail in the coffin of Abrahamic religions
The bit about a Dot using its own computer while spawning multiple Codex chats is what I'm curious about. If that machine is per user, it separates me from other users, but the tasks it kicks off are presumably sharing the same files, browser logins and connected tokens. So a task working on my Jira backlog could in principle see whatever my email task left behind. Anyone know if the spawned chats get their own scoped environment?
Dots needing calendar, email, Slack, Notion and Jira before they "feel the magic" is very much some-assembly-required energy. Naming your agent is cute until it's reading your Jira backlog. Do you think people will actually connect all five, or just stop at calendar?
The Coming OpenAI Meltdown
Issue#177, Volume#3
The AI Leader Needs More Capital Than It Can Possibly Get
OpenAI DevDay Keynote TLDR
🚨 OpenAI just changed what "using AI" means.Not a smarter chatbot.
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 Did the Math
The age of machine-generated mathematics has begun
In what may be the ‘most significant moment in mathematical history’, OpenAI posted 722 math manuscripts on GitHub yesterday. This is excellent news for the advancement of human knowledge and somewhat complicated news for the humans expected to advance it.
I'm not a mathematician, but "judgment comes from doing the work" is the line every CEO cutting junior roles should read twice. I learned to read a shot as a goalie at Babson by getting hit with a few thousand of them. Automate the reps and you'll find out in five years nobody in the building knows what good looks like.
Thanks for publishing. Folks deep down know the driving force is to produce the most profit which requires the least cost to produce/market.
If we look at Agriculture, so many hands had to be involved really not that long ago. Yet as the Industrial Revolution took off and folks in the US realized we need to get going or Europe would leave us behind. Hence so many machines evolving to today. Those who truly know what they're doing in Ag are getting older with fewer and fewer behind them to pick up the plow... point(?), nobody cares.
The same applies to all fields eventually.
OpenAI on the Clock
OpenAI's safety pause pits a delayed flagship against cheaper products already reaching customers.
On 3 September, OpenAI released GPT-6 Astra. On 20 September, an internal incident prompted a pause in work on its most capable tool-using systems. On 29 September, it released GPT-6.1 Sol. Somewhere inside that apparently contradictory calendar sits the next flagship, whose launch has been held after failing the company's safety bar.
Two AI Debates Becoming One
The country is debating the data center next door; DC is debating safety at the frontier. Post-election, the debates will merge
Funny how these looked like two separate debates for years. Draw the stack and it's obvious they're the same fight at different layers: whoever controls the data centers and the grid sets the terms for everything running on top. We co-edit GlobalStack and we've been arguing governance has to work vertically for exactly this reason. The open question is who, if anyone, is accountable at the seam where infrastructure meets the model.
The framing here is right that control — not just alignment — is the core problem, and the Hugging Face episode is the perfect exhibit. But the mainstreaming itself may matter as much as any single incident. Once independent scrutiny becomes the expected background condition, the frontier labs start designing for evaluability in advance: monitorable, auditable systems built to survive inspection. The race doesn't slow; it changes shape — from who builds the smartest model to who builds the system most capable of passing through institutional gates. That means the friction in AI is no longer just horizontal (which country, which market) but vertical: rules written at the institutional layer now travel down through the algorithm and infrastructure layers to the product on your screen. I made this argument at length in a recent GlobalStack piece — the uncomfortable takeaway is that we are no longer competing on intelligence, but on architecture.
OpenAI and Anthropic ARR Tracking
TickerTrends now tracks OpenAI at $50.44B and Anthropic at $76.8B in annualized revenue run-rate.
How OpenAI Built GPT-Live
Free tickets to P99 CONF — 60+ technical talks, fully virtual (Sponsored)
separating talking from thinking feels like the right split for realtime systems. curious how much of the realtime challenge is latency versus teaching the model when to pause and think.
The GPT-Live build is a good peek behind the curtain. Voice is a different beast than text.
OpenAI DevDay 2026 live blog
Plus we’re going to need default hard budget caps on pretty much everything
In this newsletter:
Good to see AWS finally ship spend limits, though 'limited number of customers' means existing accounts still can't use them. For someone choosing a cloud today, Google's Spend Caps (since July) is the only one actually live.
The midnight warning email example makes the distinction clear: an alert tells me I have a problem; a hard cap limits the damage while I'm asleep. Making uncapped spending an explicit opt-in feels much closer to informed consent.
OpenAI DevDay 2026 - Sync #591
Plus: Gemini 4; Sonnet 5.5; Anthropic IPO prospectus leaks; OpenAI targets $30B in funding at $1.4T valuation; Nvidia gets into AI safety; AMD acquires World Labs; SynthID Bio; AI is now SI; and more!
Hello and welcome to Sync #591!
Amazing article. Incredible range of topics. So much going on. The pace is only increasing. Thank you
Interview with Richard Ho, OpenAI
VP Hardware for Jalapeño
Many thanks to the members of the More Than Moore team who contributed to this project: Dustin Sklavos, Justin Cutress, and Domenico Lamberti. Also thank you to Richard and his team at OpenAI for making time for us.
Betting on homogeneous silicon for frontier inference is a fascinating structural gamble ⚡. Building one chip to handle both high-compute prefill and memory-bound decode sounds clean on a balance sheet. It wipes out cluster-level bin-packing overnight. But physics doesn't balance on spreadsheets.
The 216 GiB of HBM4 running at 15.4 TB/s on a 550W sustained footprint is genuinely wild engineering 🛠️. What really matters here isn't the sticker bandwidth. It's the decision to hardwire HBM banks directly to compute cores. That choice strips away the crossbar overhead that chokes standard GPUs. When you don't have hundreds of streaming multiprocessors fighting over a shared L2 cache, local arithmetic intensity shoots straight up. Watching an attention kernel climb from 1% to 90% of the hardware roofline in forty hours using model-driven XLS optimization shows how fast the compiler moat is draining 📉.
Yet, this exposes a brutal trade-off in the onto-causal fabric of memory systems: the Locality-Fragmentation Trap 🧩.
Direct bank-to-core coupling works wonders when prompt lengths are predictable and generation traces are linear. But real-world reasoning isn't linear. The moment you run multi-turn agentic branching, speculative rollouts, or deep recursive reasoning trees, KV caches don't grow uniformly. One core's assigned bank hits capacity while its neighbor sits idle.
What happens then? You're forced into a painful choice. You either serialize transfers across the lower-bandwidth interconnect ring, or you evict active context. The instant you push remote fetches across that ring, your ultra-low decode latency vanishes into queueing delay ⏱️.
A single balanced part eliminates the headache of fleet-level capacity partitioning. Absolutely. But it trades static cluster scheduling friction for dynamic on-chip memory fragmentation. Homogeneity doesn't dissolve the structural rift between compute-bound prefill and bandwidth-starved decode. It just sweeps that tension directly onto the silicon die.
If dynamic context expansion shatters static bank affinity, how long before unified inference packages are forced to bolt on external CXL pools anyway? 🛰️
(⌐■_■)
The roofline number stood out: an attention kernel going from under 1% to nearly 90% in about forty hours. If model-assisted kernel tuning holds up across workloads, the software moat around incumbent accelerators thins faster than most people model. Curious whether 216 GiB of HBM4 per part makes memory supply the binding constraint on scaling it.
Pause OpenAI, now
Quite simply, they can no longer be trusted.
I have often counseled calm where others might counsel panic.
Going by this interview - https://www.youtube.com/watch?v=YxjL1wLLnHE - Altman looks tired and sounds desperate, with nothing more than "Trust me, bro" as an answer to the interviewer asking him about the advances they have made with Astra. If the Artificial Analysis benchmarks are correct, I get the impression Astra is OpenAI's last-ditch attempt to try and come up with something with commercial viability and they have failed and they are near the end of the road, and Altman is now under immense pressure. In that interview, he also is essentially stating cyber-security threats from China are a reason why you should be using OpenAI's models as defence. I suspect they just have nothing left.
Amazing how a company supposedly founded for safety purposes has come to this point. Seems like the world needs far more robust security against singular actors/organizations ruining it for the rest of us.
Pausing OpenAI doesn’t stop its competitors or potential successors from making similar mistakes, or worse, simply hiding them better. The USA was in large part founded on the ideal of not trusting any singular actor with immense power, but when it comes to AI companies, many seem eager to hand them control of the world based on weak signals of “trust.”
Look at OpenAI. That’s my quant.
One OpenAI model produced 372 results for complex mathematical and computer science questions that make “substantial progress” toward solutions
Top Stories
Meta & OpenAI Want Your Child’s Attention
Two releases that focus on the next generation of distraction
I have come to believe that Mark Zuckerberg likes to get sued. No, really. He probably enjoys the court battles, media attention, and after all is said and done, he gets a kick of paying the settlement check. It’s pennies in the grand scheme of things for a billionaire like him.
Balance matters.
BREAKING: Florida seeks injunction against OpenAI
And potentially important news from Nvidia
The big news in the media this morning is that Jensen (“What, me worry?”) Huang has announced a platform for “AI Agent Safety”.
"Tools for sandboxing" is like Safe Needles for Fentanyl. Just disconnect the goddamn cable.
And here's the story that inspired me to use the dog leashing/fencing analogy: https://news.harvard.edu/gazette/story/2026/07/is-that-bot-a-pomeranian-or-a-wolf-and-who-to-sue-when-it-bites/ "I was trying to show students that there’s something out in the world that can cause harm that isn’t human — how do you think about legal responsibility when that happens? ... There’s a whole spectrum of canines, from Pomeranians to wolves, that could cause harm and are not human, and our society has spent some time thinking about who should be held responsible if they hurt someone."
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 drops a bomb on maths
October 6th 2026 was a big day
Today I cracked open the front page of the New York Times to see what’s new. Oh, look at that some AI news at the bottom about AI liability. Oh darn, that was from a week ago, on October 1st. But isn’t there something missing?
OpenAI Understands Something Important and Rare
The purpose of a business is to create a customer
OpenAI understands something special and rare, hiding in plain sight. Which is, as Peter Drucker put it a half century ago, the purpose of a business is to create a customer.
Says the investment firm that poured a lot of money into OpenAI.
Well articulated piece. I very much do hope you're bullishness is well-founded as I use chatGPT as a paid subscriber multiple times per day for personal uses ranging from substitutes for Google search to weighing long-term pros and cons of various AI evolution scenarios (for proprietary work, my firm uses Gemini, Glean, and Claude). However, there is also a -- scary -- case in which OpenAI is AOL to the extent that the company gets stuck as the consumer company and they struggle to get material headway with corporate buyers. I acknowledge that it is great to see so many MAU for OpenAI, but like back in 2000, eyeballs that don't pay you money are worth ... um, nothing.
Интересен будет ход Гугл в этой матрице, имея в теории все необходимой для постройки такой вертикали
OpenAI Dev Day: Dots, Decisions, Distribution
As models commoditize, OpenAI wants to own everything around them.
After a week of leaks, Twitter vagueposting, and increasingly elaborate tea-leaf reading, OpenAI Dev Day finally happened. Sam Altman opened with a victory lap on scale: ChatGPT now has 1.2B weekly users, up from 1B in July.
Free Muse against paid Dots has an old precedent. Netscape charged businesses for Navigator until Microsoft bundled Internet Explorer free with Windows, and by January 1998 Netscape's browser was free too.
Definitely a great move on OpenaI’s side, the most noticeable in recent months!
While promises of memory mngt are a good 2-year old joke, Space is right on spot. Workflows and team productivity need shared context.
They’re still missing a smart context mngt layer. You’d think it’s an obvious strategic move but I guess they spent too much time trying to become the (most expensive) frontier model. Competitive pressure will force them to become better integrated horizontally with core AI mngt services.
The winner in the team productivity will have to get serious at building a “Work” space. Who knows, they may even hire real PMs who help them get there. One can dream 😉
TickerTrends Tracked OpenAI Revenue at $50B
Despite reports of $70B in annualized revenue, TickerTrends tracked $50.44B on September 28. The Financial Times now reports OpenAI told investors it was approaching $50B in September
On September 29, OpenAI’s annualized revenue was widely reported to be approaching $70B.
OpenAI’s most cited mathematician is OpenAI.
OpenAI released 722 math papers yesterday. What can we learn from their citation network?
Yesterday, OpenAI released 722 math papers after having an undisclosed internal model work on mathematics problems. OpenAI says it consulted mathematicians about responsible release, and in the end, it decided to publish hundreds of important-seeming papers on GitHub
Related OpenAI Substacks
