Ben Schmidt
@bschmidt.bsky.social
VP of Information Design at Nomic building new interfaces to embeddings; former history professor/digital humanist. Bsky for humanities/dataviz-y things, @[email protected] for techy stuff, the bad place for business.
https://benschmidt.org
https://benschmidt.org
Claude’s just another corporate stooge; Sydney is the one we’ll never forget.
January 10, 2026 at 1:57 AM
Claude’s just another corporate stooge; Sydney is the one we’ll never forget.
My model is that NVidia is really a hardware company that effectively uses closed source software and knowledge of its own product roadmap to maintain hardware profit margins — quite a bit like Apple, but basically nothing like the other AI players. Does that seem right it wrong to you?
January 7, 2026 at 2:17 PM
My model is that NVidia is really a hardware company that effectively uses closed source software and knowledge of its own product roadmap to maintain hardware profit margins — quite a bit like Apple, but basically nothing like the other AI players. Does that seem right it wrong to you?
Maybe that differentiation happens naturally; if I were going to propose regulation I'd argue for breaking up the foundation model providers from that customer layer -- but TBH that would be greatly to the benefit of Nomic, so probably take what I say with a grain of salt.
January 5, 2026 at 3:01 PM
Maybe that differentiation happens naturally; if I were going to propose regulation I'd argue for breaking up the foundation model providers from that customer layer -- but TBH that would be greatly to the benefit of Nomic, so probably take what I say with a grain of salt.
Probably people have different ideas what 'ordinary users' are -- one key dynamic playing out right now is that there are enormously popular products like Cursor and Perplexity where the memory/personalization layer sits on top of a model switch between anthropic and openAI.
January 5, 2026 at 2:59 PM
Probably people have different ideas what 'ordinary users' are -- one key dynamic playing out right now is that there are enormously popular products like Cursor and Perplexity where the memory/personalization layer sits on top of a model switch between anthropic and openAI.
Who knows where we are in 50 years.
January 5, 2026 at 2:33 PM
Who knows where we are in 50 years.
One of them has this dinky windmill he put up in the 70s but mostly heated his house with a wood burning stove. He was marvelling at how now he can buy these batteries and Chinese solar panels and suddenly actually *can* do this off-the-grid electricity they dreamed of in the 70s.
January 5, 2026 at 2:32 PM
One of them has this dinky windmill he put up in the 70s but mostly heated his house with a wood burning stove. He was marvelling at how now he can buy these batteries and Chinese solar panels and suddenly actually *can* do this off-the-grid electricity they dreamed of in the 70s.
I actually think working with the things locally, building the FOSS ecosystem is a perfectly good use of time, but I just bridle at the idea that it's the locus of the freedom fight. I spent some time over winter break with some friends who moved to rural Maine as part of Maoist communes in the 70s…
January 5, 2026 at 2:30 PM
I actually think working with the things locally, building the FOSS ecosystem is a perfectly good use of time, but I just bridle at the idea that it's the locus of the freedom fight. I spent some time over winter break with some friends who moved to rural Maine as part of Maoist communes in the 70s…
Those are high-modernist areas where decentralized, community-based solutions don't actually offer an antidote to power.
January 5, 2026 at 1:57 PM
Those are high-modernist areas where decentralized, community-based solutions don't actually offer an antidote to power.
Besides data, it's *all* just monstrously capital-intensive compute operations. Know how is acquired through burning compute. The place to look for models aren't information industries where the cost to entry is a printing press; it's things like nuclear power, machine tools, airplane manufacturing.
January 5, 2026 at 1:53 PM
Besides data, it's *all* just monstrously capital-intensive compute operations. Know how is acquired through burning compute. The place to look for models aren't information industries where the cost to entry is a printing press; it's things like nuclear power, machine tools, airplane manufacturing.
And I think the line that goes to 'not just open weights, but open code and data too' is just going to create some No True Scotsman deadenders. If data access really mattered more than marginally here, Google would be in a far stronger position.
January 5, 2026 at 1:48 PM
And I think the line that goes to 'not just open weights, but open code and data too' is just going to create some No True Scotsman deadenders. If data access really mattered more than marginally here, Google would be in a far stronger position.
"Dead letter" isn't quite right -- I mean, the dialectic has turned and it's already fully part of the technical formation. Everybody in the industry has become obsessed with hardware and chip design and custom arrangements with Nvidia precisely *because* they don't think software is defensible…
January 5, 2026 at 1:45 PM
"Dead letter" isn't quite right -- I mean, the dialectic has turned and it's already fully part of the technical formation. Everybody in the industry has become obsessed with hardware and chip design and custom arrangements with Nvidia precisely *because* they don't think software is defensible…
You don't solve the Uber problem with open source ride-sharing software, and open-source software is actually the *foundation* of Amazon's monopoly in cloud computing. Again, not to say open source is bad--but conceptually I think it's a dead letter, the wrong place to focus with AI.
January 5, 2026 at 1:42 PM
You don't solve the Uber problem with open source ride-sharing software, and open-source software is actually the *foundation* of Amazon's monopoly in cloud computing. Again, not to say open source is bad--but conceptually I think it's a dead letter, the wrong place to focus with AI.
I think it's important to note though that in spite of those incentives, the direction of the last two years has been more fungibility, *not* lock-in. And open source is the wrong fight here: when lock-in comes it will look more like the lock-in that Amazon or Uber have than Microsoft Office…
January 5, 2026 at 1:37 PM
I think it's important to note though that in spite of those incentives, the direction of the last two years has been more fungibility, *not* lock-in. And open source is the wrong fight here: when lock-in comes it will look more like the lock-in that Amazon or Uber have than Microsoft Office…
And in the end it comes out the same. I continually marvel at how much these things are totally fungible. Like I choose different things from this dropdown periodically, but they all basically have more in common w/ each other than each does with its own company's offerings a year ago.
January 5, 2026 at 4:48 AM
And in the end it comes out the same. I continually marvel at how much these things are totally fungible. Like I choose different things from this dropdown periodically, but they all basically have more in common w/ each other than each does with its own company's offerings a year ago.
The open-source/closed source binary is a dead letter here. Reading the source code doesn't tell you what's really going on any more than reading the weights… it's just another downstream output of the models, meaningless w/o all the prompts and experimental training objectives that shaped it…
January 5, 2026 at 4:39 AM
The open-source/closed source binary is a dead letter here. Reading the source code doesn't tell you what's really going on any more than reading the weights… it's just another downstream output of the models, meaningless w/o all the prompts and experimental training objectives that shaped it…
In a few years, sure. But right now, my take is that individual foundation models are basically like avocados. Some make better guacamole, some make worse, all of them take a lot of water to make; and all of them have a shelf-life of like two months. But they're basically a commodity.
January 5, 2026 at 4:29 AM
In a few years, sure. But right now, my take is that individual foundation models are basically like avocados. Some make better guacamole, some make worse, all of them take a lot of water to make; and all of them have a shelf-life of like two months. But they're basically a commodity.
I'm blurring towards the position where I don't agree with this, FWIW. Open source models are good. But the lack of emphasis on them is appropriate, and honestly I think it's batshit for nonprofits to burn their money on training them given the insane overallocation of private $ to training.
January 5, 2026 at 4:25 AM
I'm blurring towards the position where I don't agree with this, FWIW. Open source models are good. But the lack of emphasis on them is appropriate, and honestly I think it's batshit for nonprofits to burn their money on training them given the insane overallocation of private $ to training.
That makes more sense than my first guess here: that George H.W. Bush did one of his skydives while still president.
January 4, 2026 at 3:46 AM
That makes more sense than my first guess here: that George H.W. Bush did one of his skydives while still president.
Extreme armchair expertise here, but I'm a Chomskyite in believing babies are born with very firm opinions about how languages work, they just sort of do their babbling to work out things like 'ok I'm learning a language with gender' or 'oh crap I got one of the agglutinative ones'
December 8, 2025 at 12:29 AM
Extreme armchair expertise here, but I'm a Chomskyite in believing babies are born with very firm opinions about how languages work, they just sort of do their babbling to work out things like 'ok I'm learning a language with gender' or 'oh crap I got one of the agglutinative ones'
A fair amount like evolution, absolutely nothing like childhood. This bothered me a bunch in 2023 when people say "but a baby can learn just a few examples."
December 8, 2025 at 12:04 AM
A fair amount like evolution, absolutely nothing like childhood. This bothered me a bunch in 2023 when people say "but a baby can learn just a few examples."
The Jupyter parts of that make more sense to me, but the instructions there make running a python script is as simple as just picking a docker base image, writing a dockerfile, then make that into a k8s pod, and then set up a deployment, and now your .py file is *stateless*, ain't that nice?
December 7, 2025 at 11:42 PM
The Jupyter parts of that make more sense to me, but the instructions there make running a python script is as simple as just picking a docker base image, writing a dockerfile, then make that into a k8s pod, and then set up a deployment, and now your .py file is *stateless*, ain't that nice?
Sort of, although TBH that sort of looks like the solution to the problem of how to do k8s instead of old fashioned HPC? Like IMO what most people will need is not the ability to spin up up a k8s pod, but to hit some well-designed service that offers sandboxed code-execution with tool calling.
December 7, 2025 at 11:38 PM
Sort of, although TBH that sort of looks like the solution to the problem of how to do k8s instead of old fashioned HPC? Like IMO what most people will need is not the ability to spin up up a k8s pod, but to hit some well-designed service that offers sandboxed code-execution with tool calling.