The Ledger
Every statement that passed quotation and attribution checks. Mix any filter with any other: certainty 1/5, debate potential 5/5, or both at once.
why aren't all 42 resolved? a statement only gets an assessment when the public record can support or contradict it. opinions and what-ifs never can, and 0 checkable ones are still open, waiting for their date. predictions held up or didn't; assertions are supported or contradicted. on every card: ▮▮▮▮▮ certainty · ▮▮▮▮▮ debate potential. speakers are clickable
Ian Webster: Anthropic Claude censorship matches DeepSeek on Chinese political topics
“Sure. Yeah. So, I mean, the level of censorship here is like Anthropic Claude is actually on par with DeepSeq.”
Casado: DeepSeek R1 standalone will not have a deep impact
“I don't think that R-One itself as a standalone is going to have that deep of an impact.”
Midha: Claims of a multi-year US AI lead over China were false
“No empirical evidence at the time that this was, that AI was creating net new marginal risks, and just factual inaccuracies that we were ahead of China, because if you just paid attention to what's happening, DeepSeek had already started to publish a fantastic…”
Acharya: China has sufficient compute to export models like DeepSeek globally
“A stable equilibrium is certainly not one where we just turn our back on a bunch of allies, because China's definitely has enough of the compute resources to try to export great models like DeepSeek to the rest of the world.”
Acharya: DeepSeek's rapid release disproved claims China was years behind
“Just a year before DeepSea came out, you had a number of, like, tech leaders in Washington testifying that China was, like, five to six years behind the US, with confidence, on the record. And then DeepSea comes out 26 days after OpenAI puts out the frontier.”
Webster: DeepSeek performs 20% worse than GPT on jailbreak benchmarks
“On our benchmarks, it performs about 20% worse.”
Webster: DeepSeek's safety is on par with early GPT-3.5 from 2023
“Qualitatively, what we see is performance on par with GPT 3.5, which is to say, you know, in 2023, when open AI launched GPT, there were there were a bunch of like zero day, really simple jailbreaks and deep seek is essentially susceptible to all of those.”
Webster: DeepSeek hard-censors 85% of politically sensitive topics in Promptfoo benchmark
“We did a benchmark on Chinese politically sensitive topics that, that found that about 85% of those topics in our test set were hard censored.”
Webster: DeepSeek returns CCP propaganda or refusals on Tiananmen Square
“If you ask it about Tiananmen Square or whatever, it will either give you a refusal or it will give you like this long diatribe of The CCP party line, like, you know, nothing happened. We believe in harmony in China and blah, blah, blah.”
Webster: DeepSeek is especially vulnerable to basic prompt jailbreaks
“If I had to deploy DeepSeq, I would probably focus on use cases that were not end user facing because DeepSeq is like, especially susceptible to basic jailbreaks.”
Webster: Self-hosting DeepSeek does not remove built-in political censorship
“If you run it locally or if you use any of these US providers, which have, you know, spun it up and are serving it you get the same level of censorship. The only difference there is that the China hosted version has an additional guardrail that looks at output…”
Casado: DeepSeek aggregated existing public research rather than inventing core techniques
“All of the contributions that they've done have been in the public literature somewhere. Just nobody really aggregated.”
Casado: DeepSeek's $6M reasoning model cost aligns with OpenAI and Anthropic
“The fact that they spent six million dollars just in the chain of thought is actually not out of whack with what Anthropik has now said they've spent in OpenAI has said that they spent.”
Casado: DeepSeek's success is legitimate engineering, not a PSYOP
“So I, you know, I happen to be able to believe it sounds like you are too, that this did not come out of nowhere. It's not a PSYOP. It's not, it's like, this is a great team taking advantage of what it has closely”
Sinofsky: DeepSeek's efficiency leap is AI's JavaScript moment
“And it's, to me, this is all the equivalent of the browser getting JavaScript. Because once the browser got JavaScript, then all of a sudden you could do anything you needed without going to some standards body or building your own browser.”
Casado: DeepSeek release is not a crisis for OpenAI, Anthropic, or Nvidia
“I personally don't think this is a crisis moment for open AI or Anthropic. I think like apps are hard to build. I think that like right now the apps that they put out are very complex. They actually know their users. They have very specific use cases. And so, …”
Sinofsky: The Chinese internet is a far superior AI training dataset
“Well, even on the data, you know, their starting point is, you know, the Chinese internet per se, those, that has much more structure to it. It's a much better training set.”
Casado: Shorting Nvidia over DeepSeek's launch is the wrong answer
“My reaction was not, oh, shit, I need to, like, short NVIDIA or whatever. I think that's actually the wrong answer.”
Casado: DeepSeek is a step toward AGI in your pocket
“This is another step to basically AGI in your pocket.”
Patel: NVIDIA GB200 yields 3x-4x performance per dollar on DeepSeek
“If you're running DeepSeq inference, the performance difference per GPU is like north of like six, seven X, and it continues to optimize you know, for DeepSeq inference. And so the, you know, then it's like, well, I'm only paying 60% more for six X, and it's l…”
Bosworth: DeepSeek achieved breakthrough AI results through memory architecture innovation
“DeepSeq in China, which was put in a tough spot and then innovated incredibly in the memory architectures and a couple other places to really get amazing results.”
Acharya: DeepSeek offered reasoning models for free vs OpenAI's paid tier
“So, one I think important nuance is the fact that they released their reasoning model for free at scale, so previously you had to use O-one Pro, and you had to pay ChatGPT's premium subscription to get access to it.”
Appenzeller: All future state-of-the-art AI models will use reasoning techniques
“And I think looking forward, From now on, pretty much any state of the art model will use some of those techniques, and we've seen this already, you know, from models from OpenAI and models from Google that are structurally very, very similar, and this has hug…”
Appenzeller: Generating human reasoning traces for AI training would cost $60 million
“You know, for some of the math problems, you definitely want somebody with a graduate degree. If, I mean, we saw that the average answer was 20 pages. How much do I have to pay a math PhD to generate 20 pages of text, right, if this is a I don't know, a hundre…”
Webster: Equivalent open-source reasoning models will arrive within weeks
“What I've been telling most people who ask is like, let's just wait a few weeks and there will be an open source model that, that implements this reinforcement learning technique and you, you'll, you'll get great reasoning on par with what we see from deep see…”
Webster: Enterprises should delay deploying DeepSeek until stable alternatives emerge
“The excitement around it is well warranted, but I think in an enterprise or infrastructure context, I would probably wait for something that is more stable and then doesn't have these questions hanging over it.”
Ian Webster: DeepSeek uses a separate system for political censorship
“For Deep Seek specifically, there was the part that limited speech about politically sensitive topics in China. So this is stuff like, you know, Taiwan or Tiananmen Square, that kind of thing. And it's pretty clear that that was Basically a separate system fro…”
Ian Webster: OpenAI's GPT refuses ~40% of sensitive Chinese political prompts
“Yeah, but still around 40% as opposed to 85% on this particular test set.”
Casado: China offers unbeatable global arbitrage for high-quality AI data annotation
“If you want to look at a place to arbitrage really smart, educated people and relatively low cost, it's hard to beat China globally, right? And so they definitely have access to a bunch of, you know, potentially highly educated annotated data, which is very re…”
Sinofsky: DeepSeek DAUs reached roughly 35% of OpenAI's total
“DeepSeq, I think is, was the number I saw this morning is like 35% of the DAUs of open AI.”
Casado: DeepSeek's Threat to OpenAI Is Limited if AI Apps Require Proprietary Models
“It could be that the apps actually require you to own the model. And in that case, DeepSeek is less relevant because they're not building apps. And then, you know, this means that the impact to the opening AIs or Anthropics are not as great, right?”
Sinofsky: $1T market selloff over DeepSeek was a complete overreaction
“Like on a Friday, the whole weekend was like everybody whipping themselves into a frenzy so they could wake up Monday morning and trade away a trillion dollars of market cap, which seems to be a, Complete overreaction and craziness, but that's not what we're h…”
Schmidt: ChatGPT, Llama, and DeepSeek use Nous Research's YaRN context extension
“Bone here is the lead author of a method we developed called YARN, which is a context window extension method that we released and did the research on. It is now used by every, every model you use nowadays, everything, everything Chachipiti, Lama, DeepSeq, all…”
Acharya: Real-time chain-of-thought display became an industry standard after DeepSeek
“It showed you its thought process in real time. It was just super captivating, and now something that's become a step that every model takes.”
Moore: DeepSeek mobile app achieves 7% 30-day retention versus ChatGPT's 9%
“On a retention basis, like, how many users are coming back to the app, say, Exactly 30 days, exactly seven days, exactly 60 days. It's just slightly below ChatGPT. So we're looking at seven percent day 30 for DeepSeq and nine percent day 30 for ChatGPT.”
Moore: DeepSeek reached #2 AI web product within a week of launch
“We only had 10 days of data for Deep Seek for January, because it launched at the end of the month, and it shot up from literally nothing to number two on the list, you know, 10% of ChatGPT scale on web within a week, a little bit more than a week.”
Webster: DeepSeek is slow, verbose, and outputs random Chinese characters
“Deep, deep seek isn't really a great daily driver. It's very slow. It's verbose and you know, it like throws random Chinese characters in, in its answers and stuff, stuff like that.”
Moore: Russia Is DeepSeek's Second-Largest Market After China
“So Russia is the number two market for DeepSeek after, after China.”
Moore: Runway, Otter, and UMax dropped from Top 50 GenAI list
“There were a couple companies on the list, like Runway, Otter, UMax, across web and mobile, that have been on the core top 50 ranks in the past. But maybe they got just edged out by, like, Deep Seek launching this time, and so they lost their spot for this ran…”
Mascorro: DeepSeek consistently open-sources its model weights and training techniques
“So, one of the good things about DeepSeek is basically they open source their weights, their techniques, and how they build these models, and they've been doing that for a while.”
Appenzeller: DeepSeek spent $5.5 million at market rates to train V3
“They quoted 5.5 million dollars, I think, you know, at market rates to train it.”
Mascorro: DeepSeek-R1 post-training used two SFT and two RL phases
“So basically the way they did that, trying to fix R one zero, is it added a couple more phases in the post-training. That included two supervised fine tuning phases and two reinforcement learning phases. And these reinforcement learning phases, they were a lar…”