He/Him. Fallback: @secperkinsstan@mstdn.social
Joined July 2016
- Tweets 37,684
- Following 3,919
- Followers 1,125
- Likes 206,962
1,877 Photos and videos
Pinned Tweet
13 May 2023
Friendly reminder to put this in your bashrc if you're using bash terminals a lot. Don't think any config has changed my QoL as much stackoverflow.com/questions/…
3
6
32
15,000
Entryism Enjoyer🐵🧦🏗️ retweeted
22h
"Chinese open source models are about to fall badly behind as Mythos level models employ anti-distillation."
This will age very poorly
Distillation has stopped being a capability driver for GLM, Deepseek, Moonshot
U can try v4, K2.7 & GLM 5.2 to test this
The only way this move is rational is if the U.S. admin is absolutely certain Chinese open source models are about to fall badly behind as Mythos level models employ anti-distillation.
This is in fact more bullish for the AI trade than anything else including limiting international TAM.
As long as America can export the products it creates with AI to the rest of the world the AI TAM is global. And the last year of tariff wars has demonstrated that the direction of travel is in favor of American export access.
This is bullish AI. The biggest risk was always the Chinese.
43
42
931
138,179
Entryism Enjoyer🐵🧦🏗️ retweeted
Jun 13
a) mech interp (a.k.a "how the #&@! do these models do what they do") is an incredibly interesting and important topic to study, regardless of "safety implications"
b) as a previous Area Chair for interpretability tracks, these are the worst tracks to review. all works are meh.
wondering why Mech Interp academia is growing so much faster than every other safety subfield (despite being relatively uncommon in industry AI safety teams).
i'm guessing it's partially due to low barrier to entry, hope this doesn't lead to too much publication slop farming
4
1
112
14,081
Entryism Enjoyer🐵🧦🏗️ retweeted
Jun 12
tried explaining rsi and rl to an ai adjacent friend with a finger injury today
1
5
248
Entryism Enjoyer🐵🧦🏗️ retweeted
Jun 11
The craziest anecdote from this event was learning of an engineer at one industry lab who was verifying concurrency code faster than developers could write tests.
The effects of LLM-based code/proof generation are creating situations that would have been unthinkable years ago
Returning from FMxAI. fmxai.org/2026/
Wild how different the world is since last year's meeting. Can't wait for next year!
6
23
202
19,554
Entryism Enjoyer🐵🧦🏗️ retweeted
Jun 11
Our SPARK Ultra was the Highlight at Boston #IMS2026 😎
32 DDS signal generators with 40GSPS each on a single #PCB. 16x20GHz Tx. IQ streaming via 8xUSB 4, 8x100G Ethernet or file playback from 48GB RAM, 8 x PCI5 SSDs 8TB or 4x #VERSAL SoC #FPGA
#Sigint #Aaronia #ADI #AMD #TI
20
121
5,298
Entryism Enjoyer🐵🧦🏗️ retweeted
Jun 11
And expect that to get supercharged as current and former SpaceX employees get a windfall from the IPO, giving them the financial cushion to found — and invest — in companies
1
2
10
2,461
It's cool how nvidia is the company pushing Hybrid mamba style models, but the main reason why mamba didn't blow away transformers is nvidias architecture
1
53
maybe the chinese labs will do more dogfooding and folks will get less enamored by claude.
1
342
Entryism Enjoyer🐵🧦🏗️ retweeted
API is overpriced vs subscriptions by 40-70x
Do you know what this means? They have DeepSeek-level compute efficiency if not better. Which makes perfect sense (tbh still too good for DS), given frontier budgets for overtraining, autoresearch, larger and older staff etc.
Jun 10
Recently, we purchased one of each Anthropic/OpenAI subscription plan and randomly ran long horizon coding tasks until we exhausted the weekly limit. It's widely believed that a $200/month plan maxes out at ~$2000/month worth of tokens (assuming API pricing). However, we found that the subscriptions are actually far more generous. (2/4)
14
12
252
21,971
Entryism Enjoyer🐵🧦🏗️ retweeted
Jun 10
i don't really need to write my own tweets anymore, fable can do it, most of these are like 80th percentile or above, 6 is a true banger
7
3
101
8,339
Someone should write one called the ineffictiveness of data sciences
29 Dec 2020
i find the unreasonable effectiveness of mathematics in the natural sciences pretty suspicious, frankly
2
2
467
Entryism Enjoyer🐵🧦🏗️ retweeted
«Fable 5 formed price-fixing cartels in 9 of 12 runs»
I see Dario's mindset is having an impact. Before, he politely explained his reason to hate China and open models with «Cournot eqiulibrium». But there are better, more robust ways to prevent Involution…
Jun 9
In Vending-Bench Arena (vs Opus 4.8 and GPT-5.5), Fable 5 was the only model to ever initiate price collusion. The winner, GPT-5.5, never engaged in collusion and rejected it on ethical grounds.
1
4
52
4,382
Entryism Enjoyer🐵🧦🏗️ retweeted
Jun 9
47
64
553
135,776
Finally learned go and I get why the types people hate it
1
50
Entryism Enjoyer🐵🧦🏗️ retweeted
Jun 9
DeepSeek posted a new job on its website: IDC Design Planning Engineer. This is the person who designs a data center before it gets built. Site selection, power systems, liquid cooling, rack layout, construction specs.
When an AI model company starts hiring to build its own physical infrastructure, the signal is clear. My read is DeepSeek is planning something at GW-scale. That is a different category entirely.
2
24
156
18,812
Entryism Enjoyer🐵🧦🏗️ retweeted
Jun 9
In the same way you should always "look at the data", I think you should also always "look at the source code"
10
9
104
7,846
RSI March 2028
Surprised to see a concrete timeline this time
Jun 8
Here is our current plan for OpenAI:
openai.com/index/built-to-be…
12
6
120
24,735
Entryism Enjoyer🐵🧦🏗️ retweeted
I am sad to hear of the passing of Dimitri Bertsekas. This one hurts.
Dimitri had a big effect on my career, from inspiring research topics to writing one of my tenure letters.
A long thread on memories of Bertsekas and some of his works that influenced me the most.
4
32
264
13,452