sam lessin πŸ΄β€β˜ οΈ (@lessin) header
sam lessin πŸ΄β€β˜ οΈ (@lessin) profile picture

Intern @theinformation. GP @slow. Fmr VP Product FB. co-founder dropio & fin. Seed investor solana, venmo, openphone, teamshares, sublime. https://nitter.cf/t.co/N6JqjBvF1o ΒΆ.

San Francisco, CA
Joined March 2007
Photo by sam lessin πŸ΄β€β˜ οΈ (@lessin)Photo by sam lessin πŸ΄β€β˜ οΈ (@lessin)Photo by sam lessin πŸ΄β€β˜ οΈ (@lessin)Photo by sam lessin πŸ΄β€β˜ οΈ (@lessin)Photo by sam lessin πŸ΄β€β˜ οΈ (@lessin)Photo by sam lessin πŸ΄β€β˜ οΈ (@lessin)Photo by sam lessin πŸ΄β€β˜ οΈ (@lessin)Photo by sam lessin πŸ΄β€β˜ οΈ (@lessin)Photo by sam lessin πŸ΄β€β˜ οΈ (@lessin)Photo by sam lessin πŸ΄β€β˜ οΈ (@lessin)
A bunch of people have written me back saying this was the best newsletter I have ever sent (flattering) ... so here it is for those who don't subscribe: AI Is Not a Labor Crisis. It Is a Meaning Crisis.
74
164
38
1,484
354,499
So stoked to have AZ on the judging committee for the Lessin Fellowship. more soon :)
Completely agree with @lessin's thesis that great pranksters grow into great founders, and I'm stoked to be one of the judges for the Lessin Fellowship lessinfellowship.com/
4
1
1
18
4,906
The foundation farms team looking sharp. Time to go back to business formal.
2
1
22
7,002
Lol... outlook > FAA > banks... More 'AI hack fireworks' / show of force for Anthropic IPO
1
1
24
6,422
sam lessin πŸ΄β€β˜ οΈ retweeted
yada yada yada
78
358
101
1,884
5,269,142
Watching Jensen / NVIDIA unfold in practice around a "key man" ... i wonder how many serious spacex and tesla invesors are now shopping for jumbo Elon life insurance policies -- if you broker those DM me.
6
1
21
6,133
Both outlook and the FAA?... how do I know this is AI's fault :)
7
1
23
7,580
Taped this week's More or Less Wednesday β€” mid-scandal, as the Jensen Huang / Nvidia story was still breaking in real time. A few things worth arguing about: Dave: this might be one of the best trades ever β€” NVIDIA's earnings stay monstrous no matter who's in the chair, because token demand isn't slowing down. Jessica: the board doesn't actually care about the affair β€” it cares whether Jensen lied to them, and separately whether Nvidia turned a blind eye to banned chips leaking into China. Brit: already has the Succession-style movie pitched β€” Alice, the affairs, the kids Madison and Spencer stepping in to run the company. My own take: none of this was ever really about GPUs. It's human nature at $400B of stakes β€” and proof of the thing people keep underestimating: the world isn't getting more entertaining because algorithms got better, it's getting more entertaining because everything, including the news, is just content now β€” and we consume all of it exactly the same way.
2
11
4,899
I find my bot's daily vision report from monitoring cameras oddly poetic ... what the house sees ...
5
13
5,476
lol who sent this to me?
1
8
4,295
This is very cool use of AI...
I repurposed our data pipeline at Hamlet, which is normally used to analyze public meetings, into a way to analyze all of Jensen Huang's long form interviews since 2021 (that I could find). Really interesting stuff - there are definite trends over time across a wide variety of subjects ranging from AI Safety to how much he talks about the United States, and competition. Enjoy!
1
7
5,486
let it be know that while unpublished, I called Tim Cook for NVIDIA yesterday recording More or Less … fun discussion out tomorrow
2
1
14
6,308
sam lessin πŸ΄β€β˜ οΈ retweeted
This is an absolute masterpiece. 10/10. No notes
651
8,235
781
50,269
1,892,153
Profumo 2.0 ... history rhymes, nothing is new - just bigger
1
8
4,363
"Slutty chinese spy" wold have for sure been number one Halloween costume ... near miss on timing, so close!!
2
48
7,280
sam lessin πŸ΄β€β˜ οΈ retweeted
KV cache compression is one of the most important and competitive areas of AI infrastructure as it determines how many concurrent inference sessions your GPUs can serve. Pearl Research now has the most efficient KV cache quantization on earth, all the way down to the CUDA kernels! We took a genuinely different, lattice-based approach to this problem, wasting 30% fewer bits than Google's TurboQuant and reducing native FP8 KV memory by 65% on GLM5.3.
Each inference session has its own KV cache, so a smaller cache means more concurrent users per GPU. Introducing PearlQuant, our new KV cache quantization technique: 65% less storage than FP8, better throughput and faster serving, same quality. GLM-5.3 on 8Γ—B200, 128 concurrent sessions: ⚑ p90 TTFT: 144s β†’ 7s πŸ“ˆ 24% higher peak throughput 🎯 Quality on par with FP8 pearlresearch.ai/research/bl…
25
55
8
581
35,206
Tomorrow's more or less -- which was recorded yesterday -- is going to turn out to be a very interesting real time history piece
5
3
44
12,424
Oh torah bot... u funny.
3
1
26
6,348