Ayush Goyal (@goyalayus) header
Ayush Goyal (@goyalayus) profile picture

depending on the year you are reading this in, either I will be or I am currently taking the most high stakes decisions about AGI in this new world

Joined February 2023
Photo by Ayush Goyal (@goyalayus)Photo by Ayush Goyal (@goyalayus)Photo by Ayush Goyal (@goyalayus)Photo by Ayush Goyal (@goyalayus)Photo by Ayush Goyal (@goyalayus)Photo by Ayush Goyal (@goyalayus)Photo by Ayush Goyal (@goyalayus)Photo by Ayush Goyal (@goyalayus)Photo by Ayush Goyal (@goyalayus)Photo by Ayush Goyal (@goyalayus)
If you are someone who is into post-training models over unsloth, we just made your life 10x easier.
4
11
1
93
11,477
please explicitly state the rate limits, I will buy a lot of thi s
F*CK IT $10 for $500 GLM 5.3 till end of the week
151
interestingly Astra did weight space programming way before it was famous but it was to reward hack
10
290
if you are wondering why are you not getting notifications in your mac, its because codex is using computeruse and mac blocks notifications during it for some reason.
2
180
Ayush Goyal retweeted
System design is mostly asking yourself: what if instead of bad, we simply made it good?
50
32
6
1,087
23,789
I used to work here boys
After months of building, we're finally live on Product Hunt. Do check it out: producthunt.com/products/rec…
2
176
my brain is running at 100% MMA Utilizaiton
95
Replying to @archgen_
@archgen_ is building skyscrapers for semiconductor chips Shrinking transistors is getting harder. Stacking them makes heat harder to manage. We’re building the tools to design chips in three dimensions. We stack transistors vertically, shorten connections, and optimize timing and heat together. Our foundational model for thermal simulation, predicts temperature roughly 5,000× faster than conventional EDA tools. Our placement engine ranked #1 in HRT’s macro-placement challenge We’re working with some of the largest fabs in the world to tape out our first chip in April 2027 archgenlabs.com
121
119
47
580
109,378
6.1 terra is a shit model, anybody who tells you otherwise is not looking at what he is doing, blindly trusting the model
2
242
nahh what really happens is you loose the energy to do anything else and get really focused
i have a theory that a lot of high intelligence individuals perform better with less sleep because they're generally very neurotic and a lack of sleep helps them enter a delusional state that gives them the positive attitude they need to take advantage of their abilities
120
Ayush Goyal retweeted
there is nothing complex here. if you are a hard boiled user, you should already have: > 2-3 autoresearch goals > existing project with PRs where ai was mostly ineffective earlier > existing image gen/visualisation requests > existing article writing session this would give you 90% contamination-free "feel" of the model in 12 hours or so. remaining 10% in the following week where you notice what is unique about a model that is not in your mapping.
Replying to @tokenbender
Btw im a bit curious, whenever there’s a new model release(which is a pretty frequent thing), how do you sort of assess that model’s capabilities or how much better it is compared to what we currently have?
4
4
74
4,262
Ayush Goyal retweeted
the most cost efficient SFT/RL trick that I can share with anyone is "loss to zero". one example, loss to zero, <5 USD . most issues with your pipeline, intuition for if the model can learn from data/env or not, self-doubts would be taken care of. afterwards, do it for 10.
4
2
74
4,722
somebody ban vertisium on hackertube
5
142
meri aatma trapt hogyi
You can now book our humanoid cleaners on a recurring subscription, for up to 3 hours per cleaning at tau-robotics.com. Video is first take, uncut, shown at 4× speed. See the reply for detailed footage at 1× speed.
3
237
the fact that I have to still tell claude to use tma
1
184
I have finally moved back to my dear Sol 5.6, never refuses any request, runs for long, he wants me to win!!
3
200
me : Astra go do that work Astra : okay i will do that work me : you are dumb or something? go do the work??? Astra : I am really sorry, I will just go and do the work FOLKS AT OPENAI, ARE YOU FUCKING DUMB? DON'T YOU USE YOUR OWN FUCKING PRODUCT??
1
9
753
the state of learning with LLM's is so inefficient right now that it took me 2 hours to learn how Dflash works with Claude, when it should have been a tweet how is it different from speculative decoding? 1. instead of generating one token, it generates 8 tokens at a time (you can control this but its generally 8) 2. it does not do the full prefill pass as spec decd to generate the context history's kv cache then how does it get contexts KV cache if it doesn't do its own prefill run 1. when the main model runs the prefill we save the activations of some layers (suppose n layers), then concate them, then pass them through a linear layer to reduce the dimensions (making it equal to a single dimesnion), so for each context token we get a single vector 2. our draft model also has to be n layers, now for each layer that vector we had gets dotted with K_W and K_V of that draft models every layers K_W and K_V and from that we get KV cache for all the context tokens what about kv cache of tokens we are just about to produce? suppose we are generating 8 tokens parallel and the last token approved by the main model was "I", now the 8 tokens are special tokens known as masked tokens, so this draft token doesn't have any embedding table, its just a single embedding for masked tokens. so basically KV for these tokens are this token dotted with K_W and K_V too Attention is bidirectional in the draft model generally while training transformers we mask the attention to prevent it from cheating but here none of these 8 tokens are present so it has to predict all of them so we don't mask them, infact we actually want them to look ahead The final token selection is a bit unusual too till the part where we choose the top-k tokens for each slot (masked token) is same as normal llms, things change after that. suppose we choose 4 slotsn for demonstration here the initial token is "I" slot 1 slot 2 slot 3 slot 4 cand 1: live (2) in (2) New (3) York (2) cand 2: love (2) you (2) Los (3) Angeles (2) now each of these slot tokens have two embeddings which are trained while the training of this draft model is going on, lets call one "left" and other "right" We walk left to right, one slot at a time. Each slot only checks against the word just picked before it. Slot 1 (left word = the anchor I): I → live: A[I]·B[live] = [1,0]·[1,0] = 1 → 2 + 1 = 3 ✓ I → love: A[I]·B[love] = [1,0]·[0,1] = 0 → 2 + 0 = 2 pick: live Slot 2 (left word = live): live → in: [0,1]·[0,1] = 1 → 2 + 1 = 3 ✓ live → you: [0,1]·[1,0] = 0 → 2 + 0 = 2 pick: in you can notice that embedding of "live" in step 1 is different from that in step 2, that was the left and right embedding I was talking about
2
1
13
970
lfmao
It's never too late give up on your dreams and become a data company
3
197
after a lot of introspecting I have come to realize that the perfect wiffee is always 1-2k followers on twitter
7
299