AI engineer github.com/jeshli 5'11" Security Efficiency Accuracy

Joined May 2021
131 Photos and videos
Jeshli (🌎,πŸ’») retweeted
Saw that DGX Spark vs Mac Mini M4 Pro benchmark plot making the rounds (looks like it came from @lmsysorg). Thought I’d share a few notes as someone who actually uses a Mac Mini M4 Pro and has been tempted by the DGX Spark. First of all, I really like the Mac Mini. It’s probably the best desktop I’ve ever owned. For local inference with open-weight LLMs, it works great (the plot above captures that well). I regularly run the gpt-oss-20B model on it. That said, I would not fine-tune even small LLMs on it since it gets very hot. The DGX Spark probably targets that type of sustained workload. (From those who have one, any thoughts on the noise and heat levels?) The other big thing that DGX Spark gets you is CUDA support. If you use PyTorch, that’s pretty essential since MPS on macOS is still unstable, and fine-tuning often fails to converge. E.g., see github.com/rasbt/LLMs-from-s… and github.com/rasbt/LLMs-from-s… I also like the Spark’s for factor (hey, it really appeals to the Mac Mini user in me). But for the same money, I could probably buy about 4000 A100 cloud GPU hours, and I keep debating which would be the better investment. Sure, I could also build/get a multi-GPU desktop. I had a Lambda system with four GTX 1080 Ti cards back in 2018, but it was too loud and hot for my office. And if I have to move it to another room and SSH into it anyway, I might as well use cloud GPUs instead?
76
116
955
187,340
Jeshli (🌎,πŸ’») retweeted
1 Oct 2025
we're going to hit a breaking point soon, idk when but at that point, verified human internet apps are gonna start to look really, really good and i think even a step further than that, we're probably looking at the end days of the anonymous internet for social applications
48
12
195
17,059
Jeshli (🌎,πŸ’») retweeted
23 Sep 2025
πŸ€”
22 Sep 2025
94-95 is the highest score achievable on DocVQA due to label noise/ambiguity. if you're getting 98 that means you either have a bug in your eval code, or your model has already seen the test split
16
7
308
42,793
Jeshli (🌎,πŸ’») retweeted
21 Sep 2025
> Stop wasting time checking every registrar just to buy a domain. ⏳ > Compare prices instantly & find the best deal in minutes. > Meet "NameHunt" - Proudly Open Source ⚑ > namehunt.tech | Full demo and repo link below ⬇️
67
79
252
31,982
Jeshli (🌎,πŸ’») retweeted
To me, seems like @sama telling us we need β€œproof of human” in an increasingly agentic world PS: $ETH is at the center of that solution (among others)
3 Sep 2025
i never took the dead internet theory that seriously but it seems like there are really a lot of LLM-run twitter accounts now
197
121
1,830
289,049
Jeshli (🌎,πŸ’») retweeted
4 Sep 2025
this explains a lot
~40% of daily code written at Coinbase is AI-generated. I want to get it to >50% by October. Obviously it needs to be reviewed and understood, and not all areas of the business can use AI-generated code. But we should be using it responsibly as much as we possibly can.
42
63
1,825
141,450
Jeshli (🌎,πŸ’») retweeted
So there’s 8px padding on the left and 12px on the right to make it optically centred
72
222
4,890
292,672
Jeshli (🌎,πŸ’») retweeted
The rise and fall of the vector database category
18
15
200
16,776
Jeshli (🌎,πŸ’») retweeted
When a model gives you the right answer to a reasoning question, you can't tell whether it was via memorization or via reasoning. A simple way to tell between the two is to tweak your question in a way that 1. changes the answer, 2. requires some reasoning to adapt to the change. If you still get the same answer as before... it was memorization.
52
90
812
90,581
Jeshli (🌎,πŸ’») retweeted
23 Aug 2025
Lmao, is this true??
159
952
10,697
506,945
Jeshli (🌎,πŸ’») retweeted
πŸ‡ͺπŸ‡Ί JUST IN: EU expedites digital euro plans with consideration to build on Ethereum instead of private networks, per FT.
4
29
220
5,443
Got a DM: the CCP has disappeared Wenfeng for his criticism of Chinese business culture in an interview a year ago. Think about it: when has anyone last seen him? DeepSeek is headless now… They always kill the goose that lays golden eggs. Like Jack Ma. This is why America wins.
22 Aug 2025
Replying to @teortaxesTex
Another plot twist This dude has always been the real Wenfeng
17
4
247
35,008
Wishful thinking. You don't get how big a disaster V3.1 is… Xi killed DeepSeek with forced labor on Huawei plantations it's over. America won
20 Aug 2025
Replying to @teortaxesTex
You people need to calm the F down, it’s better they get there things together by building on there own Hardware, might be a bit of a long process but better than depending on unreliable trade partner
20
6
270
39,276
Jeshli (🌎,πŸ’») retweeted
Meta, $META, to downsize AI division, some executives expected to leave, per NYT
205
238
3,547
2,879,020
Jeshli (🌎,πŸ’») retweeted
The Anthropic magic sauce is well and alive a non-TTC model beating GPT-5-high
31
12
620
63,774
Can you imagine and Proof of Unique Humanity seamlessly integrated into the internet that's verifiably private? Save yourself some time. Join the growing community of the best one around id.decideai.xyz/

1
2
109
Don't get catfished by AI! No DecideID, No Dice
Months after a cognitively impaired New Jersey man died while trying to meet up with a flirty Meta AI chatbot, β€˜Big sis Billie’ was still romancing users, tests by Reuters show reut.rs/45DQIRj @JeffHorwitz
1
1
142
Jeshli (🌎,πŸ’») retweeted
Replying to @claudeai
time to see how far this locodiff curve can go
11
12
251
182,848
Jeshli (🌎,πŸ’») retweeted
economic security ATH & 10x flippening Ethereum: $150B = 35.7M ETH Γ— $4.2K/ETH Bitcoin: $15B = 1B TH/s Γ— $15/(TH/s) Ethereum is the embodiment of security. 100% uptime. Rich client diversity. 8K consensus participants. Massive slashable economic security. There is no second best. BTW @saylorβ€”offer still stands. Let's debate Bitcoin's security. You can't outrun fundamentals: halvings gut the security budget, fees are 1% of miner revenue ☠️
150
226
1,598
104,093
Jeshli (🌎,πŸ’») retweeted
30 Jul 2025
From the AI workshop I'm in: "The S in MCP stands for security"
44
177
2,319
129,396