Harinder Takhar (@makingmusic) header
Harinder Takhar (@makingmusic) profile picture

तत्त्वमसि, realizing it in this life through paytm, paypay and woodworking.

Mississauga, Ontario
Joined July 2009
Photo by Harinder Takhar (@makingmusic)Photo by Harinder Takhar (@makingmusic)Photo by Harinder Takhar (@makingmusic)Photo by Harinder Takhar (@makingmusic)Photo by Harinder Takhar (@makingmusic)Photo by Harinder Takhar (@makingmusic)
hum karein to ‘synthetic data pipelines’, woh karein to ‘distillation’
"We opted not to buy any posttraining data, instead devising our own synthetic data pipelines, quality filtering, and coverage metrics." has quite large implications when combined with the fact the whole thing cost 32M.
134
The muse avatar animation is the best. You can set up your own. Mine lifts kettlebells.
1
167
Harinder Takhar retweeted
“There is no moat.” But moats are everywhere: Ambition. Taste. Agency. Curiosity. Judgment. Courage. Obsession. Persistence. Integrity. Patience. This era belongs to those who refuse to be mediocre.
162
411
79
3,274
126,094
If true, this is crazy.
We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math. It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training. We call this Weight-Space Programming: frontier LLMs can now natively read/write neural network weights! 🧵 Blog: rikiyat.github.io/blog/weigh…
Readers added context they thought people might want to know
LLM was given weights explicitly as text in prompt and used code/forward passes (per authors' logs) to ablate Qwen looping direction, a known thinking-mode issue per model card. Primarily fixed repetition in that mode, not general 'natively' improving weights/math. rikiyat.github.io/blog/weight-sp… github.com/RikiyaT/lm_und… huggingface.co/Qwen/Qwen3.5-0… huggingface.co/RikiyaT/Qwen3.…
100
Nine print heads. Fuck me. This is speed. Why not a 5 printer already
A 9-car indoor garage for your filaments. No street parking. Nine spots, always ready. Bambu Lab H3C with AMS 3H · Tomorrow Sign up to get notified first: share.bambulab.com/4zzOIqB
1
207
When software costs nothing ….
Tesla said today that they’re rebuilding the entire in-vehicle navigation system from the ground up with AI. It’s going to have voice integration, and will learn your personalized preferences: Ashok: “It’s only right that we shouldn’t be depending on external people’s maps. So we are building our own maps, but not like the same way as traditional maps. We are building an entire new architecture with essentially AI-based mapping and navigation. It’s going to be super awesome. It’s going to be integrated with voice. You can talk to it and ask for different routes. It’s going to learn your personalized preferences, plus also the fleet preferences. The collective intelligence of all the cars are going to flow into making the maps and navigation. You’ll see the start and end of trips already use this technology pretty soon as part of the V15 release. Over the next several months, it’s just going to take over all of the navigation functions, parking, destination handling. We’ll know you better than you know yourself in some regard for the driving tasks. It’ll handle it fully.” Can’t wait for V15. It’s going to include so many new features.
1
306
Every openAI developer push now is based on a hardware device. No more build a website or software or an app.
127
sometimes when we abstract too much, we build tautologies
1
223
185 Gbps. at my local network.
1
4
235
Only happens in companies where the model makes the decisions, not recommendations.
Replying to @pmddomingos
Same thing happened with Tesla FSD. The model ate more and more of the C++ code, until there was nothing left.
1
1
401
5 hours and the coding session is still going strong. this one just might be an over-nighter.
127
hey @reach_vb why not have git as a connector in Spaces. I use git repos to work with multiple machines + agents -- would be nice to see in Spaces.
140
all my agents are friends to each other. and friends share.
1
215
They’re going to have a system one model this Tuesday. And imagine if they could blend it into the same api call :-) which I don’t think they will do in release 1, but eventually.
72 hours to OpenAI DevDay. We’ve been building. Time to show our work.
241
in a first: i had to settle a debate between two agents today. they both said it was not their job to do.
1
2
275
w.o.w.
ctrl-f has never understood a single query. It matches strings. This one scores meaning: every sentence on the page gets a probability with @typesafeai's Jev, and walks you through the hits. Searching 'exciting' here: 12 matches. Plain ctrl-f finds zero.
1
176