Josh Terry (@JoshTerryPlays) header
Josh Terry (@JoshTerryPlays) profile picture
Austin, TX
Joined August 2014
You can be anything you want to be, but you must be exactly what is required in order to be what you choose to be.
1
2
203
Anyone else finding themselves not using whisper flow to talk to Claude because they “just don’t feel like talking”? Also, I know it’s faster, but am I actually smarter when I’m stream of consciousness on every prompt? I’m not so sure.
1
151
Best ai advert to date.
somehow the corposlops refused to use my video
1
2
368
My biggest challenge with AI, er SI, is getting it to NOT try to code intellegience into software, but to create good software that enables agents to make intellegient decisions. Anyone found first principles guardrails for this that are useful?
1
171
He got her funding, how do ya like them apples?
Of course, that’s your contention. You’re a first-year machine-learning engineer. You just got finished readin' Attention Is All You Need, probably watched a Karpathy video too. So now you’re convinced everything is just transformers and scaling laws. That's gonna last until next month when you discover convolution, and then you're gonna be talkin' about inductive biases and locality and how CNNs were actually incredibly compute-efficient for vision. Then you're gonna read the FlashAttention paper, and suddenly everything's about IO complexity and SRAM and how FLOPs don't matter because the whole goddamn thing is memory-bandwidth bound. That'll last until somebody shows you an MoE model, and then you're gonna be in here regurgitating DeepSeek, talkin' about expert parallelism and active parameters and how dense models are economically obsolete. Then six months from now you'll write one shitty Triton kernel, look at an Nsight trace for the first time, and start telling everybody Python isn't actually the bottleneck because the GPU is asynchronous. And by next year you're gonna be standing right here explaining to me that your 400-billion-parameter model is fast because you turned on CUDA graphs.
149
Josh Terry retweeted
Of course, that’s your contention. You’re a first-year machine-learning engineer. You just got finished readin' Attention Is All You Need, probably watched a Karpathy video too. So now you’re convinced everything is just transformers and scaling laws. That's gonna last until next month when you discover convolution, and then you're gonna be talkin' about inductive biases and locality and how CNNs were actually incredibly compute-efficient for vision. Then you're gonna read the FlashAttention paper, and suddenly everything's about IO complexity and SRAM and how FLOPs don't matter because the whole goddamn thing is memory-bandwidth bound. That'll last until somebody shows you an MoE model, and then you're gonna be in here regurgitating DeepSeek, talkin' about expert parallelism and active parameters and how dense models are economically obsolete. Then six months from now you'll write one shitty Triton kernel, look at an Nsight trace for the first time, and start telling everybody Python isn't actually the bottleneck because the GPU is asynchronous. And by next year you're gonna be standing right here explaining to me that your 400-billion-parameter model is fast because you turned on CUDA graphs.
Ben Affleck reveals he writes Python, understands convolutional neural networks, worked extensively with GPUs, and used his celebrity status to get private looks at Google and OpenAI’s video models “I’ve always been kind of into computers since I was young. Then, when film started to move from analog film to digital, I became more interested in that aspect of it. The visual-effects workflow for many years has included machine learning, so I can write pretty shitty Python scripts and stuff like that. “With convolutional neural networks, which were the precursors to what the transformer can do, which is much more computation simultaneously, you would do things like look at what’s called a tensor. That’s the numerical translation of a visual image in numbers, like the batch number, the frame number and the red, green and blue values of each pixel in each frame. It’s just that simple. That numeric is called a tensor. “You’d use a convolutional neural network to identify patterns that reveal what’s called edge detection or feature extraction, which is identifying patterns well enough to know, this is where the window ledge is, so we can more easily take the green-screen image out and replace it with something. “That was familiar to me early on because, prior to Artists Equity, I had a small visual-effects company. I’ve worked with GPUs a lot too. The visual-effects guys said, ‘Hey, you should see. There are a couple: Google and this other company, OpenAI, are doing really interesting stuff with transformers in video.’ “I’ve learned that I can actually just call up and go, ‘Hey, it’s Ben Affleck. Can I come see what you’re doing?’ Sometimes people say yes, to my astonishment.”
272
1,058
139
12,101
550,879
in case you're a person who doubts themselves like me. I've asked really capable people about what they're doing with ai agents. they're all creating an agentic harness using a team of agents communicating via a ledger. If you understand everything i just said you're up to date you can stop scrolling and get back to work.
1
151
it never changes... you have other products? no, not to me.
136
I think I might finally burn through my five dollar Jev budget tonight. Been lighting that money on fire for days now.
133
Harry Potter was a jock.
1
183
Someday I will make enough money to have time to listen to the All In Podcast.
2
128
Teaching agents what to do is such a strange process. You can give it a one-line instruction for something that is known and it can work for 3 days and get you a perfect result. And you can explain for 3 hours how you want something done and it's a less-known process and the agent might struggle for ages and get you nothing. Knowing when to be brief and when to go from the ground up at a granular level is quite fascinating.
1
151
Harnesses are for optimizing what’s known, not what isn’t. The best harness when unknowns is the question, “does this make sense?” Not, “does this hit metric?” Yes, continues. No, adjust and act again. On repeat. People with metrics to start with are blessed people indeed.
139
I don't think I've ever built a useful test harness. I've been doing nothing but work with AI for the last six months and I find them to be useless unless you already have access to really good data. If you're actually building something new (I'm building my own agentic harness for video) it's kind of useless. The best actual 'test harness' I have built is the instruction... "Go do this thing, then document the output of the result, look at that result, and then ask yourself the question 'does this make sense in relation to the desired outcome' and if it doesn't then theorize on what would make it better, then go do that, and try again. Do not build a harness. Do the work yourself." I have found that agents are obsessive about building harnesses, and people on social media talk about them obsessively. And they only help you get a good grade on something we already know the answer to. It doesn't actually help agents do anything beyond the painfully obvious. Maybe I don't know anything about harnesses. But for what it's worth. That prompt has been one of the most powerful prompts I've used the most often to get good results. Which in plainer english it is just telling an agent to, 'actually try stuff, and don't be lazy by trying to get a machine you built to do your work.' Well, huh, isn't that something. Advice most of us should be taking until we actually know what we're doing.
2
3
416
I swear, every two weeks, Claude gets incredibly stupid. And this for sure has nothing to do with my blood sugar levels while speaking to it.
1
2
228
Anthropic and Open AI resetting their credit limits feels like two gas stations across the street having a price war except instead of being able to afford a slurpee I get to use it to build generational wealth.
1
1
5
244
fable credits were nice while they lasted...
1
184
I'm finding that working with coding models is validating everything I've learned about learning theory and skill development for humans. Latest example: training a chat agent. You might first have a semantic router deciding whether or not something is small talk. This is two calls to respond to something as simple as small talk. Call number one - "is this small talk?" Call number two - "Respond to the small talk." This is exactly how humans inexperienced with social environments work. They first - "Try to understand what the person is trying to say. " Then second - "Try to come up with a response." But if you're going to make that chat agent more intelligent (and less expensive to operate) you'll do one of two things. 1. Hard code deterministic responses. Which is exactly what a person who doesn't want to learn too much about social environments but wants to become fast and quick in basic networking experiences. For example, a salesperson who is going to have many shallow relationships that don't last long with many people 2. Train the model to identify small talk and respond in the same call or pass it on if it is more complex to a smarter model in the first call (do two jobs at once, semantic classification and light response). Which is exactly what a person does when they want to become socially adept and move through social environments in flow while being completely engaged.
2
2
179
There is nothing you can take from a man that is greater than what he can give freely if inspired.
2
3
194
Here you go. Hopefully this helps. All of your sales problems in one graph.
4
312