Sabitlenmiş Tweet
Neural Network in @scratch from Python, hmmm...
youtu.be/eJ1HdTZAcn4
#NNIS #NNFS #NeuralNetworks #InScratch #DeepLearning #MachineLearning
YouTube
English
Daniel Kukieła (『 ᴰʰᵃⁿ🄾🅂 』)
1K posts
Daniel Kukieła (『 ᴰʰᵃⁿ🄾🅂 』)
@daniel_kukiela
ML/coding @ruliad_ai and @sentdex, co-author of the Neural Networks from Scratch book - https://t.co/HgvHw2ObbX
127.0.0.1 Katılım Şubat 2010
227 Takip Edilen925 Takipçiler
@tszzl Smoke tests, at least for me, actually increased the success rate of implementing features on the first try - the model can spot that something does not work as intended, and correct mistakes without any interference on my side.
English
I'm trying to follow open-source models for coding and get into self-hosting them with Llama.cpp. I barely set @MiniMax_AI M2.7 with a reasonable 37 t/s, and M3 has been released - a much bigger model, only 10 t/s, so not very reasonable, but still possible to use. And what I see? @xai's GLM 5.2 weights are about to be released. This model is epic from what I heard, but it's even bigger - beyond what I could self-host. And GPU prices are still going in the wrong direction. We finally have capable models with open weights, but it's also getting harder to self-host them.
English
Interesting. I'll take a look. Thank you! I even contacted them directly previously to try to find out if we can control the screens directly (other than sending full-screen images, which is slow) and if they have any protocol for partial updates, or just straight framebuffer control, and they were not very helpful. I have things related to G1 pretty high on my todo still, including writing my own firmware (I figured out a few things for this already), but maybe now we'll have an easier way? I wonder if this refactor relates to G1s as well.
English
At the Even Realities talk at AWE
Attendee: “Hey will there be a way to port apps natively to the glasses without using the phone?”
Even CEO: “Looking forward, forget about apps man. We’re in AI world now.”
English
@Sentdex Another reason is that what you send to the internet stays on the internet. Your chats are being read by humans and used in various ways. You should be aware that, for example, your research is no longer yours once you use any of these models within it.
English
Too many people on my timeline are talking about why we need OSS AI because they can just turn off the faucet at any time.
This is missing the bigger point.
We need OSS AI because this AI company literally turned that AI model against you to psychologically mislead you.
English
@Sentdex So finally, you don't need IDLE anymore :)
English
@Sentdex @MiniMax_AI @RyanLeeMiniMax @UnslothAI Can't wait for the weights and for the technical report as well. I'm so curious what hardware does one need to run it locally compared to M2.7.
English
Dear @MiniMax_AI and @RyanLeeMiniMax could yall go ahead and push the M3 weights to HF for us? And tell us how @UnslothAI already did the quants ofc.
Thanks, yall the best!
English
@Sentdex So, I also got this one, but I really had to try several times. x.com/i/status/20621…
Daniel Kukieła (『 ᴰʰᵃⁿ🄾🅂 』)@daniel_kukiela
After some more tries, I managed to get this one, but you really have to try.
English
@daniel_kukiela i think if we draw a starting point line, we'd have to say: no XD
English
Playing with Nvidia Cosmos3 Super models for image and video generation. Here's obligatory Will Smith eating spaghetti.
First few renders were pretty boring, so I went all out on an "energetic stuffing face with spaghetti prompt" here.
That's some bottomless spaghetti.
English
After some more tries, I managed to get this one, but you really have to try.
Daniel Kukieła (『 ᴰʰᵃⁿ🄾🅂 』)@daniel_kukiela
I'm trying @NVIDIAAI's Cosmos3 Nano (can't fit the Super variant). After 7 tries with different prompts, negative prompts, couple of hyperparameters and a little bit ot luck with the seed, I got this - does this pass the "horse walking backwards" test?
English
I'm trying @NVIDIAAI's Cosmos3 Nano (can't fit the Super variant). After 7 tries with different prompts, negative prompts, couple of hyperparameters and a little bit ot luck with the seed, I got this - does this pass the "horse walking backwards" test?
English
I did not appropriately prepare for this day that is apparently already here.
English
@Sentdex And I wonder if they're going to distill 3.7 from 3.7 Max as they did with 3.6. Full models are most likely going to stay proprietary
English
@daniel_kukiela I would imagine it's not likely. Qwen3.6 max never opened the weights afaik, but the smaller variants did
English
For anyone who isn't sure, this is how you release a model and talk about the performance. Not 3-5 cherry-picked benchmarks.
Qwen@Alibaba_Qwen
Performance：Qwen3.7-Max performs strongly across benchmarks in coding agents , and improves massively in general-purpose agents. Qwen3.7-Max also demonstrates exceptional strength on the hardest reasoning benchmarks, and stands out in general capabilities and multilingualism.
English
@AixAurora @Sentdex And, at least in my opinion, ChatGPT is better at research with crawling and with math. Claude is better in putting this into code.
English
thx Minimax and Hermes.
Let's see if I can survive here and maybe even downgrade further.
lately Claude has been most helpful at reading hermes sessions when things aren't working, then producing guidance where it's going wrong and saving as md references for later
English
@Sentdex Yes. The thing is, once guided, these smaller models can just follow instructions and implement things pretty efficiently and accurately - they're really good at this already.
English
@daniel_kukiela Yeah I really think we'll find ourselves way more often running smaller local models and only pulling out these larger ones for assistance rather than always boiling the ocean for simpler stuff..
English
@bitplane @vicentesurraco @Sentdex Yes, StackOverflow if full of direct answers, but it's question-answer only. It's knowledge, not "intuition" and multi-step conversation.
And user-generated messages on your model are invaluable for future model improvements in the areas the model is less capable, makes mistakes.
English
@vicentesurraco @daniel_kukiela @Sentdex Every message you send to Claude Code is a message it's learning to send to itself. They're training autonomous agents with higher level planning and steering. We're its future inner monologue. Stack overflow doesn't have step by step task completion data, Claude Code users do.
English
Apparently an unpopular opinion, but I don't think Anthropic owes anyone heavily subsidized tokens for their third party app.
Theo - t3.gg@theo
I can't help but feel personally burned by the Claude Code changes announced today. We put so much work into wrapping the (atrocious) Claude Agent SDK in T3 Code. It was the ONLY path they supported, so we made it work. It was hell. Now our users are getting their rate limits cut by 40x, despite us doing everything right. I listened to the Claude Code team. I had my issues with their direction, but I trusted them and took them at their word. I will never make that mistake again. Until we see significant change, it is safe to assume any statement from an Anthropic employee is a lie on a timer. The rug will be pulled, no matter how many promises are made beforehand.
English
You do not have to train on this directly, you can extract the mistakes where the user clearly stated them and what the response should be (useful for RL), you can extract user reactions, expectations, tasks, etc.
I don't know which plan you are at, but it's not unusual for me to hit the limits. Also depends on the model, context size and if you're getting back to sessions that already expired in the KV cache (which means counting all the history as input tokens) - and these things are not clear about how they set them.
English
@bitplane @Sentdex @daniel_kukiela I am unconvinced. First off, AI training on AI code, even when led by a human, would likely lead to a decline in code quality.
Second, I (and many others) rarely hit anywhere near usage limits. Perhaps "maxing out" the usage is subsidized, but by who? To some extent, other users
English