I collect dots to connect them later. Views are my own. Building https://xitter.twitterapi.xyz/t.co/QIx831U6wn Writing on https://xitter.twitterapi.xyz/t.co/rOzI5yJv5t
Bengaluru, India
Joined December 2011
- Tweets7.2K
- Following619
- Followers1.5K
- Likes6.2K
Met someone who uses ai rampantly but recommends people not to use ai.
My take is that, ai should not be gatekept by people with influence if they themselves use ai to amplify their productivity. – at Bongaigaon, India
1
5
345
More benchmarks needed!
Nice DeepSWE score! Does it survive a lowercase filename?
Introducing MaxxBench: 89 DeepSWE tasks with small changes to the instructions or starting code. Same engineering problem, slightly different requirement.
Models clustered around 74% on DeepSWE end up 23 pp apart.
Good at coding, or good at DeepSWE? 🧵
3
229
This is so cool! – at Bongaigaon, India
Launching Junction: come speed-run the semiconductor tech tree this December in Bangalore.
We’re flying in students from all over India 🇮🇳 to a 2-week fabrication retreat at @_lagrangepoint to build equipment, design chips and engineer processes, with our friends at @hackerfabindia.
India still needs thousands of semicon engineers -- the ecosystem to produce them isn’t mature enough yet -- we're changing that.
Current applicants have built vacuum chambers to study plasma physics, RISC-V processors from scratch, scanning tunneling microscopes, et cetera
At junction you will:
1. Propose a project
2. Figure out a plan with the Junction team.
3. Iterate fast towards a working prototype, with help from the Junction team
4. Be mentored by top semiconductor hackers and specialists.
5. Demo to industry leaders, founders, investors, and your fellow craftspeople
Flights, accommodation and project expenses? All covered. Apply on the link below
4
230
Never ask instinct to go through your mails and create a list of your expenses, categorized.
I repeat, never do it. – at Bongaigaon, India
2
101
I wrote this 2 years back. Holds true today
blog.iamsohan.in/blog/the-fu… – at Bongaigaon, India
54
He is right in his claims, but misses the point.
The "small fix" did not involve retraining, RL Envs or some LORA math.
It simply analyzed and updated the weights to do the fix. – at Bongaigaon, India
Sorry to disappoint everyone, but GPT-6 Astra didn't teach Qwen math.
Before the edit, Qwen3.5-0.8B ran out of tokens on 93% of math answers. It was stuck in thinking loops. After the edit: 33%.
Astra's final report: of 579 practice answers it fixed, 576 had been cut off. On answers Qwen already finished, accuracy didn't move.
On Gemma 270M, which barely loops, no edit moved math by more than a point. According to the authors, editing weights "could not add an ability the model lacked."
What Astra did is still impressive. It compared what happens inside the model when it loops vs when it finishes, found the signal two of its layers write when it loops, and edited it out. Then it checked that a random edit of the same size didn't do the same.
But this isn't something people haven't done before.
57
100% true. It can come off as blunt.
Using AI well requires going meta: instead of solving a problem, design a system that solves the problem.
Using AI *really* well requires going meta-meta: build a system that continuously improves itself.
Using AI on God mode requires building a system that watches you and builds automations so that you're never doing what an AI is able to do better.
If this sounds dystopian, sorry to break it to you but you're still stuck in a pre-AI mindset. Whether you like it or not, the world has changed. So instead of operating how-things-used-to-be, operate in the reimagine-everything mode as your bandwidth gets free from the mundane.
2
187
I don't walk back this, but yes, reading and reviewing every line of code is perhaps less important.
I was never a fan of unc. I always hated clean code and this is rather hypocritical to come from him.
He used to fearmonger saying if you don't write clean code, something will fuck up and regulators will get involved and start issuing licenses for SWEs.
Not reading code risks exactly that. That means it was always about selling books and very rarely about building an engineering culture that values practicality, durability, reliability and versatility.
53
This is extremely clever, I am not sure how useful it is, but I can see some utility.
1/4
Meet Qwenputer: Qwen3.5-4B with a full RISC-V CPU inside its weights, using the model's own Gated DeltaNet layers as RAM.
It runs C and MicroPython in its forward passes, no external runtime.
On DeepMind math: base 55–90%, base + Python tool 95–100%, Qwenputer 97.5–100%.
1
84
This is unequivocally, 100% true.
The first problem with AI-generated text detectors is the assumption that 100% AI-generated means zero human thought.
A non-technical person and @karpathy could both generate 100% of their code using AI. That doesn’t mean the code they produce reflects the same level of engineering skill.
The same applies to writing.
1
1
103
If Paras runs the ministry of compute, I can more or less guarantee that effective compute per capita is immense.
India needs a ministry of compute.
Like ministry of oil, we need laser-focus from the top to build compute necessary for participating in the future economy.
Otherwise, we will repeat in a decade what we’re doing now: spending expensive forex for importing what we don’t have.
5
31
1,029
