Ujjwal

Introducing Claude Opus 4.8: it builds on Opus 4.7 with sharper judgment, more honesty about its own progress, and the ability to work independently for longer than its predecessors. Available today at the same price.

You can find all articles on my profile. My website has my Toward Data Science articles too: neuralavb.com/articles/

@akshitakauzi Same taste bro! It was more sweet

@UjjwalCodes How did you like it?

For the first time in life, i saw yellow watermelon from inside

you know what all of these "which is better" polls are silly use codex or claude code, whatever works best for you i am grateful we live in a time with such amazing tools, and grateful there is a choice

“if we ever stop talking, and you don't know how to come back, send me a song.” • Begin Again

“Everyone can code now!” Dude, no one can code now.

@neural_avb Are dhyaan mt do sir! Haathi chale bazaar, kutte bhoke hazaar!

Tired of AI replies. More tired of hate comments/rts that have suddenly increased. All from 10 followers, sus accounts. Idk what I have done to offend anybody. Whatever. This is going to be the default setting from now on. Hope you are still able to comment.

@neural_avb Finally a unsloth video before GTAVI, thank u sir once again!💕

New video just out on Finetuning SLMs! Build 0.1B models that run locally on device at ~350 tok/s and perform narrow tasks - Generate synthetatic data on general tasks with Outlines - SFT with Unsloth - Build harnesses and SDKs Find this on youtube (helps the channel!)

It’s not anyone, its my shattered dopamine receptors

Emotional Manipulating v/s Gaslighting

@neural_avb Finally unsloth yayyy!

All notes ready for my post-training video How to train SLMs to do cool shit at millisecond latency directly on client device > generate synthetic training data locally > low-level guidance > Unsloth training > evals > creating harnesses around narrow SLMs Recording tonight!

Thank you for being so real here @ThePrimeagen 🫡 He echoed so much of what I have been feeling about dealing with dying skills. It's been 5 years since I finished my MS and some skills I honed during my uni + initial work are slowly becoming obsolete too. When I was starting out, writing a sentiment classifier from scratch using LSTMs was considered impressive. Today an LLM will one-shot that shit. Think about what this does to the mind of a young dev who's starting out: - "You used AI to make a good project? Its not impressive coz everyone is using AI now - whats your moat?" - "Oh you wrote a great project on your own? Oops that was slow, somebody else shipped 5 more products during that time." "Plus we suspect you used AI anyway coz we don't trust anybody's skills right now." You are simply not being rewarded for pushing yourself to the edge for 2 months building something crazy the same way I was rewarded 5 years back, or Prime was rewarded 10 yeas back. Now that might be seen as "waste of time" (ITS NOT!) The result of all this is - the best policy for young devs has become: "Build a ton of slop and pad your resume with them. Don't waste time learning coz anything high-effort slows you down" And that is the worst worst worst outcome of all this. AI was supposed to make us more excited to learn. But the hype marketing + insane spike in intelligence means that things you would have loved is getting automated. Tangential thought: the only positive change that has come to me as a dev is: I don't take code personally anymore. Code is so cheap I can just write and delete a bunch of it without attaching any amount of self to it. I see it more and more as a commodity now and less as a reflection of my skill. The skill now is to design better software, test it, and know how to distribute it. Thats the reason somebody like Prime is thriving in this landscape. He's an established dev who also built an audience. Heck, even me - a much much smaller creator - I have been able to make a connection with my audience through making niche high-effort content. The moat is distribution and proof of skill. Remember the old skills I said was dying? Turns out even though I can't recommend a hand-rolled deep learning pipeline at an office job (coz there probably exists an API for that in 2026), I can still create good educational content around it. And people are likely to enjoy learning something niche thanks to (creators like) me. I make good money doing this. This is one of the reasons I encourage younger devs to work on content creation, writing blogs, and creating educational videos on the side. It is a proof of skill, plus it builds an audience and a network, and it is one of the few professions that is yet to be disrupted by AI. High effort content is gold today. And it is fun to make. Anyway long post. Once thank you Prime for articulating it and being real as usual.

If youre even a little bit ambitious, you will almost always be miserable not because youre not doing well, but because you expect a LOT from yourself and you're always in a hurry to do things you constantly compare yourself with others (even if you know that its completely irrelevant) wins become meh, which means even if you win you will treat it as the least acceptable possible outcome, not a reason to celebrate but a threshold that you must always pass guilt kicks in even if you take a day off or don't open your laptop, work all day everyday and you're not satisfied either weird place to be in

We engineers were CPU bound and are now IO bound. Earlier we used to 'think' while coding, now we just prompt Claude and block.

LLM Knowledge Bases Something I'm finding very useful recently: using LLMs to build personal knowledge bases for various topics of research interest. In this way, a large fraction of my recent token throughput is going less into manipulating code, and more into manipulating knowledge (stored as markdown and images). The latest LLMs are quite good at it. So: Data ingest: I index source documents (articles, papers, repos, datasets, images, etc.) into a raw/ directory, then I use an LLM to incrementally "compile" a wiki, which is just a collection of .md files in a directory structure. The wiki includes summaries of all the data in raw/, backlinks, and then it categorizes data into concepts, writes articles for them, and links them all. To convert web articles into .md files I like to use the Obsidian Web Clipper extension, and then I also use a hotkey to download all the related images to local so that my LLM can easily reference them. IDE: I use Obsidian as the IDE "frontend" where I can view the raw data, the the compiled wiki, and the derived visualizations. Important to note that the LLM writes and maintains all of the data of the wiki, I rarely touch it directly. I've played with a few Obsidian plugins to render and view data in other ways (e.g. Marp for slides). Q&A: Where things get interesting is that once your wiki is big enough (e.g. mine on some recent research is ~100 articles and ~400K words), you can ask your LLM agent all kinds of complex questions against the wiki, and it will go off, research the answers, etc. I thought I had to reach for fancy RAG, but the LLM has been pretty good about auto-maintaining index files and brief summaries of all the documents and it reads all the important related data fairly easily at this ~small scale. Output: Instead of getting answers in text/terminal, I like to have it render markdown files for me, or slide shows (Marp format), or matplotlib images, all of which I then view again in Obsidian. You can imagine many other visual output formats depending on the query. Often, I end up "filing" the outputs back into the wiki to enhance it for further queries. So my own explorations and queries always "add up" in the knowledge base. Linting: I've run some LLM "health checks" over the wiki to e.g. find inconsistent data, impute missing data (with web searchers), find interesting connections for new article candidates, etc., to incrementally clean up the wiki and enhance its overall data integrity. The LLMs are quite good at suggesting further questions to ask and look into. Extra tools: I find myself developing additional tools to process the data, e.g. I vibe coded a small and naive search engine over the wiki, which I both use directly (in a web ui), but more often I want to hand it off to an LLM via CLI as a tool for larger queries. Further explorations: As the repo grows, the natural desire is to also think about synthetic data generation + finetuning to have your LLM "know" the data in its weights instead of just context windows. TLDR: raw data from a given number of sources is collected, then compiled by an LLM into a .md wiki, then operated on by various CLIs by the LLM to do Q&A and to incrementally enhance the wiki, and all of it viewable in Obsidian. You rarely ever write or edit the wiki manually, it's the domain of the LLM. I think there is room here for an incredible new product instead of a hacky collection of scripts.

@ishhasr Cant say, bcoz i never had a good deep sleep!

@neural_avb Thank u sir! Mein yeh cpt waala dekhta hu tab tak

@UjjwalCodes Lolol. CPT video aa chuki hai. Aj SFT waala train karuga. Next week tak pakka video aa jaegi hopefully.

Crazy day - generated 100K+ rows of instruction data across 10 tasks and 100 full-length papers - macbook so fried, using it to iron my shirt - published the full local data-gen repo - tomorrow I train a SmolLM with Unsloth that was one out of the 3 projects I worked on today

I don't bring it up much on X, but the whole reason my youtube channel exists is because of my Patreon backers. We crossed 2.5K lifetime members. For the entirety of last year, we have had 100+ paid members/mo. Today 20+ also support through YT memberships. I am a very small creator and I make niche technical content around AI and Deep Learning. Videos take 2 weeks to 1 month to produce coz my brain is slow and obsessed with details. I am growing (~30k subs), and it's taking time. If my only source of income on YT was through ads, it would simply not be financially smart to make the videos I make. I am fortunate these people made this gig viable. Over 800 unique people have contributed financially to my channel since Jan 2025. Some supporters have been contributing for well over 12 months! I usually make extra content, share slides, docs, and make certain coding projects Patreon-exclusive. It's not a transaction, but more of a thank you note.