Brief and to the point
Welcome to Cautious Optimism, a newsletter on tech, business, and power.
Today’s newsletter is both late, and abbreviated. TWiST had a morning taping, so CO is all out of whack today. Still, far be it from us to stop making noise! — Alex
📈 Trending Up: Microsoft-Salesforce sniping … BOOM goes Boom’s non-booming jet … shooting our own feet … venture capital spine? … climate startups … idiocracy …
📉 Trending Down: Tech stocks, again … not interest rates … chip availability in China … truth … the Chicago school …
DeepSeek-ing OpenAI’s secret sauce
It feels like we’re on day 673 of the DeepSeek saga, but let me catch you up on the latest:
Microsoft and OpenAI were worried a while back that DeepSeek was trying to learn from their AI models.
Current US AI czar David Sacks underlined the possibility, saying that the Chinese company might have borrowed from OpenAI. In response to a question on the DeepSeek-OpenAI point, here’s what Sacks had to say (CO transcript, slightly tidied):
It’s possible. There’s a technique in AI called distillation, which you are going to hear a lot about, and it’s when one model learns from another mode. What happens is that the student model asks the parent model a lot of questions. Just like [how] a human would learn. But AIs can do this, asking millions of questions, and they can essentially mimic the reasoning process that they learn from the parent model. They can suck the knowledge out of the parent model.
There’s substantial evidence, that what DeepSeek did here, is that they distilled the knowledge out of OpenAI’s models. And I don’t think that OpenAI is very happy about this. I think one of the things you’re going to see over the next few months is leading AI companies taking steps to try and prevent distillation.
We’ll see if the leading AI companies can prevent distillation by third-party companies. That would slow down some of these copycat models.
Some folks felt that defending American AI companies — OpenAI in particular — by stating that they might have been copied smacks of cope. In non-Internet vernacular, the argument is that by being forced to cry theft in the wake of DeekSeek’s big week, domestic AI shops are showing that they are scared — that they are struggling to cope.
Those concerns were not not inflamed by Anthropic CEO Dario Amodei penning a post on what DeepSeek managed and did not. Similar cries of cope were levied at Anthropic as had been tossed at OpenAI.
What did Amodei actually say? Summarizing aggressively:
DeepSeek’s pretrained V3 model in December was good and cheap, but follows regular AI cost curves.
DeepSeek’s R1 model uses V3 while adding reinforcement learning, similar to how OpenAI’s o1 model works. Therefore, Amodei writes, R1 is “much less interesting from an innovation or engineering perspective than V3.”
We’re early in the reinforcement learning era, so folks are catching up to current leaders more quickly than perhaps some expected. Still, trends in AI development (model quality) and price (quick declines) should continue as more is wrung out of the human mind.
Amodei got dinged by some for using the above to argue that as chips will remain key tools in the AI war, export controls to China could help keep AI built in democratic nations on top. I think that that is, in fact, an incredibly reasonable argument.
Viva la open source. Viva, even, la quasi-open source. But when it comes to democratic versus authoritarian development of AI, I know where my allegiance lies. And it’s at home.
Other short notes:
No one knows what is going on with Federal spending. To be frank, it’s a mess and not one that bodes well for an administration this new that had so much time to prepare.
Meta and Microsoft earnings are coming out shortly. More on the key numbers and capex pledges in the morning.
The Fed’s decision to not cut rates today is reasonable, but tees up a conflict with the Trump administration that could get messy.
Back to regular service tomorrow!