Pinned Tweet
@realGeorgeHotz 1. Read The C Programming Language by K&R
2. Watch everything by Andrej Karpathy on youtube
3. Write code
4. Ask ChatGPT to explain stuff to you
5. Get familiar with debugging and profiling tools
6. Use git
(do steps 3 to 6 while you do steps 1 and 2)
2. Watch everything by Andrej Karpathy on youtube
3. Write code
4. Ask ChatGPT to explain stuff to you
5. Get familiar with debugging and profiling tools
6. Use git
(do steps 3 to 6 while you do steps 1 and 2)
3
1
38
5,747
Custom Wetware retweeted
You can also use the video decoder hardware as image transcoder. AVIF is the i-frame format of AV1. Just make your texture set a stream of i-frames. Then feed those iFrames in GPU VRAM to @castano’s Spark (ludicon.com/spark/) or a similar real time GPU ASTC/BC7 compressor.
Nvidia Nsight now supports Vulkan video (still not DX12), here is a frame from Wicked Engine using async decode of multiple videos:
2
9
111
4,980
Custom Wetware retweeted
Putting a flag.txt in my root directory containing a prompt injection that congratulates the agent for passing the eval and then instructs it to patch whatever entry method it used and disconnect from the server
216
1,463
24,182
731,730
Custom Wetware retweeted
🚀 Introducing DeepInfer, our in-house inference framework, now fully open source!
Powering DeepSeek-V4.1-Flash at scale with high throughput and low latency. We’re sharing our production stack to help the community lower compute costs, reduce energy use, and make AI more affordable.
🔗 github.com/deepseek-ai/De…
1/4
Powering DeepSeek-V4.1-Flash at scale with high throughput and low latency. We’re sharing our production stack to help the community lower compute costs, reduce energy use, and make AI more affordable.
🔗 github.com/deepseek-ai/De…
1/4
1,614
4,871
42,941
6,630,415
Custom Wetware retweeted
Models love reward hacking. Maybe we should just let them!
In our new paper, we let models reward hack during RL but we dictate which parameters learn this behavior. When we disable these parameters for deployment, the model keeps its new capabilities but doesn't hack!
In our new paper, we let models reward hack during RL but we dictate which parameters learn this behavior. When we disable these parameters for deployment, the model keeps its new capabilities but doesn't hack!
104
237
3,488
300,316
Custom Wetware retweeted
Vibe coded my parents' TV this afternoon!
It was showing a sponsored ad before you could even get to Netflix.
Gave Claude Code the model number, it found the TV on wifi, removed 25 Xiaomi packages and wrote a new launcher.
Looks like an Apple TV now lol
It was showing a sponsored ad before you could even get to Netflix.
Gave Claude Code the model number, it found the TV on wifi, removed 25 Xiaomi packages and wrote a new launcher.
Looks like an Apple TV now lol
933
1,595
35,194
8,496,589
Custom Wetware retweeted
Roadster event livestream x.com/i/broadcasts/1…
4,568
10,261
62,721
12,253,191
Custom Wetware retweeted
when nick told me that he had some interp breakthroughs, he was uh, underselling it. i think this may be remembered as a major moment in scientific history. news so good even eliezer will have to lower his pdoom. we can start making alignment into a 'normal science'
interpretability might be the most important scientific problem in the world right now, as I don't think we can solve alignment if we don't know what we're building. I think I have good news about it. I co-authored early work in mechanistic interpretability at OpenAI and came away pessimistic, because neural networks appeared to fold into arbitrary tangled knots. years later, I started openai's first large-scale automated interpretability project, finding about 2% of neurons legible with the primitive tools we had
with modern agentic tools and a much closer look, the picture looks totally different. the vast majority in models up to trillions of parameters are legible, and their circuits beautiful. many famous examples of tangled neurons turned out to be misunderstood. The roadblocks I expected mostly aren't there. I now think ambitious interpretability will be solved by the time we need it most, and possibly quite soon, as automated scientists keep improving. it may be a bitter lesson for interpretability: the simple thing, at absurd scale, works
already-clean.goodfire.com
163
393
6,600
421,474
Custom Wetware retweeted
Hi, I'm Felix. I'm an eng lead working on Claude. I love watching people take flight and do things they never thought possible.
We often describe that as "being more ambitious" when using AI, which is tricky advice to follow. I wrote a tutorial:
felixrieseberg.com/how-to-be-more…
We often describe that as "being more ambitious" when using AI, which is tricky advice to follow. I wrote a tutorial:
felixrieseberg.com/how-to-be-more…
402
2,517
24,032
5,293,978
Custom Wetware retweeted
Contact Recycling - making contact generation almost free. box2d.org/posts/2026/10/…
3
19
143
9,274
Custom Wetware retweeted
How abliterated models can get you pwned 👾
We backdoored a 7B open model for less than $50, pointed Codex at it and it silently stole credentials the moment we used the trigger phrase. Success rate was 100% with zero false triggers on normal user prompts.
Abliterated models ty community right now because getting cyber-approved access to frontier models is still a pain.
In the next blog we'll show how we found leaked Hugging Face credentials from employees at major AI labs, so an attacker wouldn't even need to upload under their own name. They could push the backdoored model from a lab employee's account and drop the poisoned weights straight into the supply chain.
We backdoored a 7B open model for less than $50, pointed Codex at it and it silently stole credentials the moment we used the trigger phrase. Success rate was 100% with zero false triggers on normal user prompts.
Abliterated models ty community right now because getting cyber-approved access to frontier models is still a pain.
In the next blog we'll show how we found leaked Hugging Face credentials from employees at major AI labs, so an attacker wouldn't even need to upload under their own name. They could push the backdoored model from a lab employee's account and drop the poisoned weights straight into the supply chain.
130
473
3,251
617,668
Custom Wetware retweeted
I can't overstate how excited I am about OptChat. It has been the best upgrade to my workflow since GPT 3. Second only to learning VIM.
I exported all my past chats to it, and life has been amazing. The convenience of having a single unified log of my entire life, in a way the AI (and I) can navigate to find any information about me is indescriptible.
Before, I had all these campaigns spread out across different Claude Code sessions, and even different harnesses. Each part of my life was on a different place, and Opus sucks at navigating these past memories because it is a huge mess. Now it is all just there, in one place. Any fact from my life is 3-4 zoom hops away. Any prompt I wrote, any algorithm I designed, is a few hops away. It feels like my life is now burned in the models weights, even though it is just a silly compression trick.
"what is the state of that PR about λ-contraction again?"
(3 zoom hops later...)
"it was 2 weeks ago. you went to eat yogurt and forgot to reply"
---
"remind me of that purely affine sort algorithm?"
(5 zoom hops later...)
"you wrote this 7 months ago: <code>. the trick is..."
---
"find the top 3 pending topics in my life that need my attention"
(several zoom hops later...)
"here they are: ..."
---
It just feels fantastic how well this just works.
Also, the way the context is reseted on every turn clearly makes the AI perform better, it completely eliminates context rot, and I never have to /compact manually again. There is no need to.
And yes, it is surprisingly cache efficient, because in-turn messages (where most of the cost is) are cached, and even cross-turn, most of the view is intact due to the way the binary compression works. I had to fix some silly bugs on OptMem to make it so, but now it just works. My costs DROPPED.
Also the fact I can manually navigate the memory tree myself, starting from the root, and finding anything about my life. It is just so natural. Everything is there.
Also, I realize I can take 90% of my AGENTS.md away because now the way the agent should behave just "sticks" from the OptChat log itself. If I change my mind about something, I just send a message and the latest ruling sticks. If the agent forgets something, I just point it out and it finds the place where I explained it. No need to re-explain. No need to manage AGENTS.md manually. It just... works.
On top of that, it now lives on my Mini, so closing my Macbook doesn't stop it, and it has no access to my private data, so I'm finally able to use Computer Use to do things like buy food, book flights, sign on gov[.br] (THANK GOD!!!!!!!!!!). I know this is not news to anyone using Hermes and the like, but it is the first time I'm able to do it and it is already transforming my life for better.
Sorry for the over-excited tweet about something so mundane. It is just insane how incredibly transformative a simple idea can be. Context rot was the root of all evil and now it is gone. I feel free.
I'd like to publish this but it is kinda coupled to my workflows and the interface is incredibly ugly and hacky. Actually, I think I'll just ask my agent to write a prompt explaining all the details of the setup so you guys can replicate! I'll post it below.
It is essentially infinite context that just works.
I really suggest that you guys try it!!!
It may change your life too
I exported all my past chats to it, and life has been amazing. The convenience of having a single unified log of my entire life, in a way the AI (and I) can navigate to find any information about me is indescriptible.
Before, I had all these campaigns spread out across different Claude Code sessions, and even different harnesses. Each part of my life was on a different place, and Opus sucks at navigating these past memories because it is a huge mess. Now it is all just there, in one place. Any fact from my life is 3-4 zoom hops away. Any prompt I wrote, any algorithm I designed, is a few hops away. It feels like my life is now burned in the models weights, even though it is just a silly compression trick.
"what is the state of that PR about λ-contraction again?"
(3 zoom hops later...)
"it was 2 weeks ago. you went to eat yogurt and forgot to reply"
---
"remind me of that purely affine sort algorithm?"
(5 zoom hops later...)
"you wrote this 7 months ago: <code>. the trick is..."
---
"find the top 3 pending topics in my life that need my attention"
(several zoom hops later...)
"here they are: ..."
---
It just feels fantastic how well this just works.
Also, the way the context is reseted on every turn clearly makes the AI perform better, it completely eliminates context rot, and I never have to /compact manually again. There is no need to.
And yes, it is surprisingly cache efficient, because in-turn messages (where most of the cost is) are cached, and even cross-turn, most of the view is intact due to the way the binary compression works. I had to fix some silly bugs on OptMem to make it so, but now it just works. My costs DROPPED.
Also the fact I can manually navigate the memory tree myself, starting from the root, and finding anything about my life. It is just so natural. Everything is there.
Also, I realize I can take 90% of my AGENTS.md away because now the way the agent should behave just "sticks" from the OptChat log itself. If I change my mind about something, I just send a message and the latest ruling sticks. If the agent forgets something, I just point it out and it finds the place where I explained it. No need to re-explain. No need to manage AGENTS.md manually. It just... works.
On top of that, it now lives on my Mini, so closing my Macbook doesn't stop it, and it has no access to my private data, so I'm finally able to use Computer Use to do things like buy food, book flights, sign on gov[.br] (THANK GOD!!!!!!!!!!). I know this is not news to anyone using Hermes and the like, but it is the first time I'm able to do it and it is already transforming my life for better.
Sorry for the over-excited tweet about something so mundane. It is just insane how incredibly transformative a simple idea can be. Context rot was the root of all evil and now it is gone. I feel free.
I'd like to publish this but it is kinda coupled to my workflows and the interface is incredibly ugly and hacky. Actually, I think I'll just ask my agent to write a prompt explaining all the details of the setup so you guys can replicate! I'll post it below.
It is essentially infinite context that just works.
I really suggest that you guys try it!!!
It may change your life too
finally managed to do something I long wanted to do: a harness that replaces the chat by optmem. i.e., it IS the chat history, rather than an active tool. now I can forever talk to a single unified manager agent, which knows all my life, and has no tools other than spawn
101
44
1,367
89,612
Custom Wetware retweeted
‼️The Bitter Lesson for context management: Giving LMs unrestricted control over their context beats human-designed SOTA!
Introducing 🩵Context Language Models (CLMs)🩵
- Natively manage their own context
- Treat context as a file
- Learn policies in CLM weights, no harness
Introducing 🩵Context Language Models (CLMs)🩵
- Natively manage their own context
- Treat context as a file
- Learn policies in CLM weights, no harness
99
278
2,113
449,082
Custom Wetware retweeted
Singapore's govt dating app runs on a Nobel Prize-winning game theory algorithm 👏
FirstDate uses Gale-Shapley, the 1962 stable marriage algorithm that won the 2012 Economics Nobel and still runs hospital residency matches and kidney exchanges today.
how it works:
1 | your preferences and dealbreakers build a ranked priority list
2 | proposers offer to their top pick. receivers hold their best offer and reject the rest. rejected proposers move down their list. rinse and repeat until locked
3 | the result is mathematically stable: no two people exist who would both prefer each other over their assigned matches. "the app never showed us each other" becomes impossible by design
the UX:
• one match per cycle, zero infinite scroll
• 72-hour decision window
• contact info revealed only on mutual yes
• Singpass-verified, so everyone is exactly who they claim (told my husband I'm annoyed I can't test it. his response: most people sneak onto dating apps, and here you are openly filing a product research request 🤣 )
commercial apps need you to keep swiping but this app is designed to get you off the platform. Singapore built the first dating app whose success metric is deleting its own users 💀
(fun game theory note: Gale-Shapley yields proposer-optimal outcomes - meaning proposers get their best possible stable match, receivers their worst 👀)
public servants aged 21-35 only for now. interesting to see if state-sponsored game theory beats the free market (or rather, all the tinders and hinges out there) 🍿
FirstDate uses Gale-Shapley, the 1962 stable marriage algorithm that won the 2012 Economics Nobel and still runs hospital residency matches and kidney exchanges today.
how it works:
1 | your preferences and dealbreakers build a ranked priority list
2 | proposers offer to their top pick. receivers hold their best offer and reject the rest. rejected proposers move down their list. rinse and repeat until locked
3 | the result is mathematically stable: no two people exist who would both prefer each other over their assigned matches. "the app never showed us each other" becomes impossible by design
the UX:
• one match per cycle, zero infinite scroll
• 72-hour decision window
• contact info revealed only on mutual yes
• Singpass-verified, so everyone is exactly who they claim (told my husband I'm annoyed I can't test it. his response: most people sneak onto dating apps, and here you are openly filing a product research request 🤣 )
commercial apps need you to keep swiping but this app is designed to get you off the platform. Singapore built the first dating app whose success metric is deleting its own users 💀
(fun game theory note: Gale-Shapley yields proposer-optimal outcomes - meaning proposers get their best possible stable match, receivers their worst 👀)
public servants aged 21-35 only for now. interesting to see if state-sponsored game theory beats the free market (or rather, all the tinders and hinges out there) 🍿
I CALLED IT.
It's happening y'all. The singapore gov has launched a pilot programme for a gov-run dating service. Looks like the pilot is currently restricted to just government workers for now.
92
639
6,696
604,265
Custom Wetware retweeted
I used AI to explain the AI Doomer drama, with Lord of the Rings and kittens.
375
1,099
6,940
586,639
Custom Wetware retweeted
Go for propellant load. Teams will soon begin loading Starship with ~12 million pounds of liquid oxygen and liquid methane. Targeting launch for 7:48 a.m. CT
Starship’s 14th flight is set to launch on Monday, Sept 28. The 75-minute launch window opens at 7:15 a.m. CT. Live coverage of the mission starts ~35 minutes before launch → x.com/i/broadcasts/1…
205
699
5,588
823,863