A few weeks ago DeepSeek (which has a whale for its logo) showed the world that Chinese AI companies could build great models for far less money than their Western competitors, and despite a limited supply of GPUs. While it was fun to watch this set off a huge firestorm of coverage and market volatility, it was also a bit of a bummer, because all of this attention meant that the party was over. 

And indeed it is being ruined: DeepSeek seems to be running slower (I’m back to experimenting with Claude), you can’t access its API page, and, finally, Missouri Senator Josh Hawley has proposed legislation where individuals using DeepSeek would face penalties of 20 years in prison or $1 million in fines, with much higher fines for corporations. This seems high. 

It’s unclear whether Hawley’s proposal applies to downloading the free, open LLMs that DeepSeek makes available, or reading the papers they publish. We’ll find out, I guess. Someone has already started tracking all of the global regulations against DeepSeek, and put them on a map. I’m glad we’re on this journey together.

It’s worth stating it plainly: I don’t fully trust any of the AI companies to do the right thing with anyone’s data. I have used DeepSeek to experiment with code projects and summarize web pages, but I wouldn’t upload a bunch of emails or share a personal spreadsheet with it. I have slightly more trust with ChatGPT and Claude, because ultimately they are bound by American laws, but the legal frameworks around this stuff is changing fast. I would not upload anything sensitive or important, or mess around with private, confidential customer, or user data without explicit permission and planning. 

Am I being paranoid? Sure. I’m not saying you have to run an LLM locally on a machine inside a Faraday cage with Wi-Fi chips ripped out and no internet access or Bluetooth in a sub-basement covered in tinfoil. But this is a new technology, and it’s very hard to craft a good threat model around it—and we don’t really know the actors involved, foreign or otherwise. 

Aboard is building a product that helps people build software with AI, and finding the right threat model is challenging. As Zeynep Tufekci wrote in the New York Times:

If the inevitable proliferation of A.I. endangers our cybersecurity, for example, instead of just regulating exports, it’s time to harden our networked infrastructure—which will also protect it against the ever-present threat of hacking, by random agents or hostile governments. And instead of fantasizing about how some future rogue A.I. could attack us, it’s time to start thinking clearly about how corporations and governments could use the A.I. that’s available right now to entrench their dominance, erode our rights, worsen inequality. As the technology continues to expand, who will be left behind? What rights will be threatened? Which institutions will need to be rebuilt and how? And what can we do so that this powerful technology with so much potential for good can benefit the public?

All good questions worth debating and discussing, if anyone can keep their focus long enough to do so. (At Aboard, we’re hoping our new office is a nice place in NYC to discuss these AI issues. We’re waiting for the construction crew to take out the weird tiki bar left behind by the prior tenants.)

Anyway: A few more interesting quotes that I’ve been thinking through, for your perusal. First from James O’Donnell in the MIT Technology Review:

You may have heard…that DeepSeek is energy efficient. That’s true for its training phase, but for inference, which is when you actually ask the model something and it produces an answer, it’s complicated… The problem, at least to some, is that this way of ‘thinking’ uses up a lot more electricity than the AI we’ve been used to. Though AI is responsible for a small slice of total global emissions right now, there is increasing political support to radically increase the amount of energy going toward AI. Whether or not the energy intensity of chain-of-thought models is worth it, of course, depends on what we’re using the AI for. Scientific research to cure the world’s worst diseases seems worthy. Generating AI slop? Less so. 

I’ve also been wrestling for a couple weeks with this blog post by Laurie Voss—some days I agree with it, others I don’t:

Jobs are more than collections of tasks. Jobs require prioritization, judgement of exceptional situations, the ability to communicate ad-hoc with other sources of information like colleagues or regulations, the ability to react to entirely unforeseen circumstances, and a whole lot of experience. As I said, LLMs can deal with a certain amount of ambiguity and complexity, but the less the better. Giving them a whole, human-sized job is way more ambiguity and complexity than they can handle. It’s asking them to turn text into more text. It’s not going to work.

And this from Ethan Mollick, in his newsletter, regarding OpenAI’s new “research” mode:

At the end, I get a 13 page, 3,778 word draft with six citations and a few additional references. It is, honestly, very good, even if I would have liked a few more sources. It wove together difficult and contradictory concepts, found some novel connections I wouldn’t expect, cited only high-quality sources, and was full of accurate quotations. I cannot guarantee everything is correct (though I did not see any errors) but I would have been satisfied to see something like it from a beginning PhD student. You can see the full results here…

I think there’s a tricky paradox, though: The fact that an AI can write in a “human-like” way and create a “research paper” with citations is very interesting right now, but it seems like the wrong kind of output. 

In general, I think that AI companies are so busy simulating human artifacts that they’re forgetting to make digital artifacts. (Counterpoint: When Claude writes little in-browser apps and runs them, that’s sick.) I don’t really want a research paper written in weird AI-prose, but I do want something that looks through lots of sources and finds connections, summarizes, and visualizes the output, so that I can get smarter. 

I had it write me a research paper on a subject I’m interested in (new types of music synthesis) and it was fine, but these are internet creatures and they should make things: Software, summaries, charts, and visualizations. This will take a long time to unlock because what I’m asking for is for the system to behave less predictably—while all the resources at hand are going towards making it behave more predictably. One step at a time.