내말이. 이거 다들 모르더라고.
해외에서 피클 사기 논란이 터진 이유
1. 피클은 해외에서 이미 사기꾼으로 유명한 팀인데, 한국에선 아무도 모른다고 함
2. 대표 이름을 보고 같은 사람인 걸 확인했다고 함
3. 피클은 원래 AR 안경을 만든다며 크라우드 펀딩을 받은 회사라고 함
4. 올해 초에 개당 $200씩 보증금을 받고 새 AR 글래스를 팔았다고 함
5. 그 스펙이 전부 비현실적이었고, 데모 영상도 완전히 가짜였다고 함
6. 판단은 자유지만, 내 개인정보를 어떤 사람들한테 넘기는지는 알고 쓰라고 함
제가 리서치해봤을 때에는 사기꾼이라고 단정하기는 어렵고
과장 홍보 논란이 있었고
결국 환불엔딩을 맞이했다고 보시면 됨다
378
요즘 한국에 범람하는(할) AI에이전트 서비스를 보면서 드는 생각:
00년대 한국이 자국 검색엔진 시장을 지켜냈듯이, 20년대 자국 에이전트 시장도 지켜냈으면 좋겠다. 수출도 했으면 좋겠다.
뭐 그런 생각;
1
2
79
루스 retweeted
생맥주 뚜껑에 고무줄을 감아서 잘 열리게 한 맥주집,
참기름 소스 뚜껑을 열때 아래 쪽이 잘 깨지거나 너무 눌려서 다 흘러내리는 걸 방지하기 위해 아래 부분만 2겹으로 보내오는 참치집,
모든 소스통을 비닐 랩으로 감아 최대 참사는 막겠다는 의지를 보이는 중화요리 집,
난 이런 디테일을 확인할때면 리뷰에 꼭 이걸 언급한다. 잘 안보이는, 남들은 잘 모를 디테일을 알아줄때가 정말 힘이 난다는걸 알기 때문이다.
I think we'll see V4.1F-1125 or 1210, or maybe even V4.2F, before we see V4.1 Pro or any other DeepSeek-Pro. They probably are scaling it up with their new compute.
그건 맞지
The people mocking the @BMW demo today are the same people that defended this with their lives in 2021. Don’t be a Tesla homer. Celebrate advancements in autonomy when it’s right in front of your face.
(Image stolen from an old @DirtyTesLa YouTube, because I had no idea where I could find this image other than going to the OG beta tester. Thank you)
79
현대(포티투닷)가 스타트 끊고 레거시들 줄줄이 자율주행 데모 보여주는 중
근데 BMW는 저 데모 환경에서도 개입을 하네
[BMW, 2027년부터 도심 주행 보조 확대 발표]
- 최신 시연 동영상 공개
✅ 내비 목적지를 설정하면 고속도로, 국도, 도심까지 이어서 보조 (SAE 레벨 2)
✅ 도심에서 좌우회전, 차선 변경, 교차로 양보, 회전교차로 통과 지원
✅ 대상 차종: iX3, iX4, X5, 3시리즈, 7시리즈 (OTA 지원)
✅ 유럽 새 규정(UNECE R171 DCAS 시리즈 2)에 맞춘 실도로 인증 테스트 통과, 2027년 초 승인 예정
✅ 독일에서 먼저 시작하고 다른 나라는 순차 적용
2
198
갑자기 맞는 말을 하는군
> Pricing in China has become so competitive that it pushes Chinese labs to find shortcuts, i.e., distillation
I disagree with this completely, "involution" is typical China watcher thought terminating cliche and particularly meaningless here. First, the logic of distillation is vastly simpler, American labs are far ahead and their models are not just developed with a lot more compute but make use of the high-quality data Chinese labs haven't had the opportunity to acquire yet, so using them to accelerate the progress is a natural solution. The core pressure is *international competition*, not domestic margin slashing. Claude is still massively popular and regarded in China, despite Dario trying his best to reject this income. If Chinese models fall further behind, the Chinese consumers will be utilizing Western models even more; and the only response would be to, indeed, slash margins and compete on price alone, in the normal involution fashion.
But more importantly, Chinese models are not cheap at all.
I've just completed a natural experiment, ran out of OCG $10 subscription, almost exclusively burning V4.1F – one of the most cost-efficient ways available to use the most compute-efficient Chinese model on the market. The estimated API rate would have been $76.5. It is *massively* subsidized. Indeed, it's likely below DeepSeek's own boundary of profitability. There are many ways to do this mafs but consider that they've reported one GPU printing $38/hr. In other words I've used up about 2 GPU-hours' worth.
[Another way: in that same paragraph DeepSeek reports 155T tokens/day off a single server, which as previously established has at least ≈2400 GPUs. 64B/GPU/day, 2.66B/hour. This way we get even worse results for DS, read on].
How much for 2 hours of GB200? is it more or less than $10? At rental prices it's a straight loss. But they own hardware. How does it work in the most optimistic case? A GB200 that makes you $5/hour is a GB200 that makes you $43.8K under 100% utilization over a year. Retail unit price like $65K, 1833 W/GPU in a rack, $0.080/kWh state-subsidized electricity, so $1300/year on GPU power alone, add a bunch of associated expenses, and we arrive at about 2 years to breaking even on compute, a far cry from Liang's target of 10 months.
Except GB200s in China are smuggled with an unhinged markup and easily go above $200K. If this were the main channel of revenue, it will JUST NEVER BREAK EVEN, by the time you approach net profit your hardware is obsolete.
But as I've said, it's not a story of involution. What volume has that virtual, subsidized $76.5 bought me? A paltry ≈7B "tokens", 98% cache hits, 107M in 35.3M out. Given that input is 1/4 the output for DS, we can normalize to 62M total "out". 6.2M tokens per $1 of actual payment. Back in Jan 26, @she_llac had reconstructed Anthropic's subscription pricing: on Pro plans, you were getting max 0,54M *OPUS* tokens per $1, with cache hits entirely free, ie infinitely cheaper than Whale API's $0.003/0.006 (which is why I'm ignoring them as well; OpenAI has made cache virtually free now, too, though we can quibble about API vs chat, 72h vs 12h etc). So the difference per token is about 11.5x. That's a lot, but Claude Opus tokens have way more utility. Since then, Anthropic subscription kept getting *more* efficient. What are we getting these days? $5K worth of API on $100 sub? $7000? Haven't been checking.
…I need to go so will leave the remaining mafs to those interested. My point is: American labs have bodied Chinese ones. More capability, better hardware, more money to burn result in a situation where the most efficient Chinese model *on the most subsidized plan* is very close in price per token to the best American model currently served; Kimi or Zhipu are nowhere in sight.
If you think, like I do, that DeepSeek is actually very efficient and Opus costs substantially more per token than a 8/16B active Flash tier, then it can be argued that *Americans* are doing involution, and not the Chinese.
90
😀😀
KAIST ranks #2 in the number of papers accepted to NeurIPS 2026, ahead of MIT and Stanford!
While paper count is not the primary metric we pursue, it certainly means something. A proud moment for the KAIST AI research community!
76




































