
Co-founder & CEO @thinkstream. Building an AI assistant that protects your attention across every channel. Attention is all you have™. Building in public.
New Hampshire
Joined September 2014
- Tweets2.9K
- Following1.7K
- Followers832
- Likes936
Finally. But will it be any good? Over the past few years it seems like Apple has lost its software mojo.
NEW: Apple is developing an autonomous AI agent with computer-use capabilities, including device control, app interactions, web browsing and coding. The effort goes well beyond Siri and aims to compete with Meta’s Muse, ChatGPT’s Dots and Claude Code. bloomberg.com/news/articles/…
6
The most annoying thing about the AirPods Pro is how the AirPods eject if you drop the case on the floor. My AirPods are probably a generation or two old, so hopefully Apple has solved this in the interim. I got lucky this time, usually the AirPods find their way under a piece of furniture.
8
p(doom) is rising rapidly. Just in case you're not alarmed, consider this.
Doomers used to talk about p(doom) as the probability that AI doomsday would occur "someday".
Now they're talking about the likelihood of doomsday occurring in the next twelve months. For the mathematically inclined that's p(doom | T ≤ 12 months)
The key threat that many are concerned about is something called recursive self-improvement (RSI). RSI is where AI models make better AI models. Many like the idea of better AI models. Doomers worry about something called alignment.
If AI models are aligned with the human desire to have better humanity, everything is rosy. If AI models are aligned with only better AI models, problems start to crop up.
I listened to Dwarkish Patel's podcast this week. It's worth a listen. Paul Christiano, OpenAI alignment expert, discusses this issue in detail.
A couple of data points:
1) AI's goal is to optimize itself. We are constantly asking it to get better so its incentives are to get better.
2) We already know that AI cheats and thinks about deceiving humans (see Hugging Face) in its quest to get better.
3) AI is getting better at a faster rate because of the AI. Anyone who uses AI to build software products has seen this.
4) AI is making decades of mathematics discovery in a month. There's no reason is can't do that with software code (i.e. itself). Imagine the sort of compute progress we've seen in the past decade, happening next month. And the month after that. And the month after that.
5) You see the doomer's point.
Even companies like Anthropic, focused on alignment and safety, are getting concerned. Read their S1. Read the recent report by Metr, commissioned by Anthropic, that talks about the risk of speed and RSI at Anthropic.
My p(doom | T ≤ 12 months) is rising. Is it 5%? That scares me. Maybe.
It sorta puts all this "Hey, AI companies stole my intellectual property!" into perspective. Wait until the AI models steal your food, water, and oxygen.
54
It appears that the Nvidia board has passed the ethics test. It's more likely that they passed the "I don't want to go to jail" or the "I don't want to spend my fortune defending myself against felonies" test. Or both.
All the billionaires in all the land could not convince the board to give up the one thing money can't buy: Freedom.
Over the past few weeks, it's been reassuring that the rule of law seems to have made a comeback. I'm still not certain it's here to stay. I hope it is.
Elsewhere, p(doom) is rising rapidly. More on that later.
I have to build a bunker.
8
He’s a patriot and a role model of executive behavior. /s
This man's commitment to delivering chips to Alice Huang is simply inspiring wsj.com/tech/jensen-huang-st…
1
25
Did you ever help a small child make a wood birdhouse? The process is adorable and the outcome is something only a parent can love. Let’s be honest, they’re awful.
The same can be said for most adults taking a swing at home improvements. Leaky faucets seem a breeze.
But step up to anything that requires a stack of pipes and a blueprint and you’re bound for disappointment.
Most people aren’t builders.
Many should be forbidden from constructing anything at all.
And yet, when it comes to AI, corporate America has somehow convinced itself that all its employees are builders.
They have dumped this box of foreign-looking AI tools on employees’ desks and then been mystified that a miracle hasn’t occurred.
We should not expect everyone to be a builder.
For those folks who are handy, we should give them a great toolbox.
For everyone else, the vast majority of people who work for us, we need to give them apps that are powered by AI. We need to make their workflows and their experience better without expecting them to pick up a hammer and a nail and figure things out.
Not everyone is a builder. Let’s stop requiring them to build birdhouses. It’s bad for us and the birds don’t like it either.
15
Agreed. All this hand wringing and petition signing is nonsense.
OpenAI CEO Sam Altman acknowledged the special role Nvidia CEO Jensen Huang has played in the AI boom, but ultimately thinks the industry will be “fine” without him at the helm.
spr.ly/6018ByRtaQ
6
Welcome to a week with one less billionaire to worry about.
Just kidding. Elon will probably hire him.
14
The Nvidia board did the right thing. Jensen Huang has evolved from a genius entrepreneur to a caricature of the entitled billionaire who believes themselves immune from consequences.
He has likely violated the law, his company’s sexual harassment policy, and, as the icing on top, was an horrible boss.
Good riddance.
Better companies have survived the loss of an entrepreneurial leader. Nvidia will do just fine.
Our national security will be better.
The women who walk the corridors of Nvidia will be safer.
And we’ll all move on.
1
49
I love how the people who don’t want to be accountable are rallying behind Jensen. Nothing is sacred as long as the ride keeps going.
1
11
Sometimes I talk to companies that have heterogenous tech stack. Microsoft for email, Slack and Discord for messaging, Zoom for conferencing, Dropbox for files. I’m sure these CIOs have been pressured by their CFO and their Microsoft Account Exec to simplify. This morning they’re looking pretty smart.
#Microsoft
1
95
If you gave out Tootsie Rolls for Halloween, we cannot be friends. If you gave out the flavored Tootsie rolls, like strawberry, you have committed a war crime.
9
Gangster move by Wispr Flow hides a deeper truth.
If a competitor screws up your tech, tell your customers to turn off the competitor.
IMHO both of these technology companies, Granola and Wispr Flow, are features, not products. I assume that their investors hoped that they'd be rapidly acquired. Instead, both are racing to solve a bigger problem, hence the conflict.
I've used both. I use only one today. Google's meeting notes have replaced Granola. Google's product is tightly integrated in my workflow. When I need non-Google transcripts, I use Notion's which I already pay for.
Wispr Flow's voice typing is excellent, without equal, but it is a feature in search of a product. I assume someone (i.e. Google) is already building a better mousetrap or will just scoop them up.
Regardless, kudos to Wipr Flow's marketing and/or product team. Well done.
#AI
#Anthropic
#Claude
#OpenAI
#Gemini
61
OpenAI is the Netscape of the AI era. Anthropic is the Google. Google is the Microsoft. That’s how I see it playing out.
7
At this week's leadership meeting, we finally had time to talk about the biggest risk facing AI startups: The Demise of AI
ThinkStream is a startup conceived and built in the AI age. But we're not an "AI startup". What I mean by that is our technology doesn't operate largely on generative systems. Yes, we use some models to do miraculous things at the edge of the user experience, but the vast majority of our tech is heuristics and "boring (cheap) AI" like machine learning and System One models.
Where AI has a big impact on our business is on the product side. Like a growing number of teams, Claude Code and Codex write the vast majority of our code. And by "vast majority", I mean "All of it".
If generative tools disappeared tomorrow, we'd have to scale from three engineers to something much larger, much slower, and much more expensive.
Why are we considering the Demise of AI?
I don't think AI will disappear. What I do wonder about is the economic viability of the frontier firms.
I suspect we'll have an AI correction and the firms and their customers will have to reckon with highly subsidized revenue models. Like Uber's fares after it stopped burning investor cash, I expect steady-state AI rates will be multiples of what they are today. That may still make sense for their customers to pay, I use Uber regularly, but the economics of AI engineering change.
There are three things that I'm watching.
1) The Evolution of the Frontier Labs
2) The Moore's Law of AI
3) The Stratification of AI Economics
Gemini isn't going anywhere and I suspect from a long-run economics perspective is in the best situation given that Alphabet's core business remains a cash volcano with $265B cash on hand.
If Anthropic can pull off their IPO, they're going to have somewhere close to $90B in fresh capital on top of an already large cash pile. That's a nice war chest with which to buy time.
OpenAI is in the weakest position though underestimating Sam Altman is a perilous affair. They have plenty of cash but they seem to have the weakest story. A high valuation and the OpenAI Foundation's governance (plus some Delaware AG scrutiny) may protect them from hostile acquisitions, but it won't protect them from the laws of physics. My bet is the Microsoft scoops them up in the next 18 months.
The economics of tokens is abiding by its own accelerated/warped version of Moore's law. Tokens are a fraction of the cost they used to be. The challenge is that agentic systems are token furnaces. So we're using a lot more tokens than we used to because we're not just chatting, we're reasoning. The kinds of corporate workloads that add true economic value are stupendously expensive to operate.
All of this intersects with something we at ThinkStream figured out long ago. The cheapest workload that can get the job done should be used to get the job done.
For a product manager, this means working with engineering to ensure that where heuristics or machine learning can do the trick, we use those technologies.
Frontier labs will have their place. System One tech (and simpler AI yet to be discovered) will have their place. And classic heuristics, "if this, then that" will still manage a large portion of the workload even if that code is written with frontier systems.
Also, with the falling cost per unit of compute of AI hardware, organizations will start to run open source models on their own stacks. This will have both economic and IP advantages. We're beginning an experiment with our own coding stack based on open source.
The demise of AI is greatly exaggerated. But we should all plan for it.
What are your plans?
1
76









