HNHacker News
TopNewBestAskShowJobs

ismailmaj

297 karma · joined August 9, 2022

ML Infra

[[username]][[hardy number]] at g mail

submissionscomments
ismailmaj··on MegaTrain: Full Precision Training of 100B+ Parameter LLMs on a Single GPU
Obsolete because of what? Because with limited hardware you’re never aiming for state of the art, and for fine-tuning, you don’t steer for too long anyway.
ismailmaj··on 25 Years of Eggs
I don't know why people mess with tesseract in 2026, attention-based OCRs (and more recently VLMs) outperformed any LSTM-based approach since at least 2020.

My guess is that it's the entry-point to OCR and the internet is flooded by that, just like pandas for data processing.

ismailmaj··on BitNet: Inference framework for 1-bit LLMs
Assuming 2 bit per values (first bit is sign and second bit is value).

actv = A[_:1] & B[_:1]

sign = A[_:0] ^ B[_:0]

dot = pop_count(actv & !sign) - pop_count(actv & sign)

It can probably be made more efficient by taking a column-first format.

Since we are in CPU land, we mostly deal with dot products that match the cache size, I don't assume we have a tiled matmul instruction which is unlikely to support this weird 1-bit format.

ismailmaj··on Surpassing vLLM with a Generated Inference Stack
Confusing, since this is specific to an architecture that no one making money will use (8B is consumer space, not enterprise). The produced code shouldn't hold much interesting IP?
ismailmaj··on Why the global elite gave up on spelling and grammar
Oh no, cortisol spike in my text-only forum.
ismailmaj··on BitNet: 100B Param 1-Bit model for local CPUs
The title and the repo uses 1-bit when it means 1.58 bits tertiary values, it doesn't change any of my arguments (still xors and pop_counts).
ismailmaj··on BitNet: 100B Param 1-Bit model for local CPUs
You drop the memory throughput requirements because of the packed representation of bits so an FMA can become the bottleneck, and you bypass the problem of needing to upscale the bits to whatever FP the FMA instruction needs.

typically for 1-bit matmul, you can get away with xors and pop_counts which should have a better throughput profile than FMA when taking into account the SIMD nature of the inputs/outputs.

ismailmaj··on Yann LeCun raises $1B to build AI that understands the physical world
Ah yes, OpenAI the puppet of Microsoft that is currently declaring war against GitHub, sounds logical.
ismailmaj··on Surpassing vLLM with a Generated Inference Stack
Any place we can find the code?
ismailmaj··on Yann LeCun raises $1B to build AI that understands the physical world
Don’t think that’s a fair interpretation of what I said.

Liquid money rich? No.

Can get pulled for big tech packages? Also no, for most of the employees.

AFAIK, big tech didn’t aggressively poach OpenAI-like talent, they did spend 10M+ pay packages but it was for a select few research scientists. Some folks left and came but it boiled down to culture mostly.

ismailmaj··on Yann LeCun raises $1B to build AI that understands the physical world
I don't see a world where they become threatening and the employees don't become rich from investors flooding in.
ismailmaj··on I put my whole life into a single database
In my experience, tracking objective things like "nutrition" and "sleep hours" is immensely useful to reflect on what went wrong, and tracking subjective things like "mood" or "stress" is useless given hedonic adaptation or heavy swings that make problems obvious, and not need tracking.

What's key is be able to visualize metrics easily on the data and frictionless data entry, I've got a decent setup with iPhone Action + Obsidian + QuickAdd scripts on Obsidian Sync (mobile + laptop). for visualization I use Obsidian Bases and Obsidian notes that run Dataview code blocks and Chart.js, couldn't be happier.

I could track things that are not interesting to reflect on like vitamin D supplementation for accountability but I've never bothered, especially if it's taken ~daily.

ismailmaj··on My “grand vision” for Rust
Surprised with the push back of the comments, getting effects on Rust would be a dream.

Could even enable some stuff like passing loggers around not by parameters but by effect.

ismailmaj··on Tech employment now significantly worse than the 2008 or 2020 recessions
Don’t need the EU for that, they are hitting everything in 2026 including unemployment, though nothing passed yet.
ismailmaj··on MacBook Air with M5
I bought the M4 air given that the consensus was also that it was the best value for 1k, but I ended up returning it and went for pro base model for a few reasons (still valid for M5 AFAIK from a bit of research):

- pro motion (120hz screen).

- better display brightness which is important when there is a bright sun outside.

- 1 more USB-C port and HDMI port (no dongle hell).

- 20% more battery life.

- This is more personal, but 13" is too small and 15" is too big, so 14" MBP worked best for me (~25 HFOV with a stand + KBM).

It's hard to justify saving 400 bucks given the gap between the models, but the decision is closer since the air has 16GB memory by default since M4 AFAIK.

ismailmaj··on Show HN: I built a sub-500ms latency voice agent from scratch
Semantic end of turn being 300ms of silence is horrible because I ended up intentionally um-ing to finish my thoughts before getting answer.

It was difficult to detrain and that made me stop using voice chat with LLMs all together.

ismailmaj··on Gemini 3.1 Pro
3.1 feels to me like 3.0 but that takes a long time to think, it didn't feel like a leap in raw intelligence like 2.5 pro was.
ismailmaj··on YouTube Is Down
Cannot trust those damn Americans during on call.
ismailmaj··on Async/Await on the GPU
Warp specialization is an abomination that should be killed and I'm glad this could be an alternative.

I hope they can minimize the bookkeeping costs because I don't see it gain traction in AI if it hurts big kernels performance.

ismailmaj··on Gemini 3 Deep Think
top 10 elo in codeforces is pretty absurd
ismailmaj··on Ask HN: What are you working on? (February 2026)
What I'm getting from you is that, given competence (and here it's a manager being able to make himself dispensable) things "work out".

I tend to agree but I have a feeling that it takes a certain number of bad apples min-maxing KPIs to destroy the system.

When the manager evaluates you based on how many lines of code you submitted because the VP decided that the teams with the most code submitted deserve a bigger bonus, then you stop doing the rational thing and start playing the game, it may be possible for the middle manager to play around a bad VP but from an IC level it's rarely worth it, you might invest your small political capital only to be burned afterwards.

I've seen way too often in my career benevolent IC ignoring the politics, getting things done only to be last to be recognized when it's clear things would be falling apart without them. Even worse, doing the right thing might keep/hide the dysfunctional system for longer.

Sorry for being overly negative, I like the theory but failure points while rare are common enough for me to have seen them a few times, if things are not rotten it probably holds water.

ismailmaj··on Ask HN: What are you working on? (February 2026)
I see your point, this reminded me of the incompetent/malicious section.

Yes, leadership could do the right thing, but the right thing is difficult to measure by shareholders and can only grasped by shareholders until it is too late (Disney for example), but why would they?

I agree, people are not malicious but the system currently selects for people that ignore the difficult to measure right thing to do (in SWE it’s doing the grunt platform work that no one will appreciate) and skip to things that are rewarded (cool feature tm that users hate by gives a cool slide to the CEO), it might seem like people can do both but it’s not in practice due to having limited energy.

Also stream of thought answer, but I feel like we are back to the trust thing, the right thing to do is impossible to measure, so we can only hope they do it because they care about doing the right thing, but incentives punishes them and makes them scarce at the top and the only solution is to go back to a high trust society (maybe like Japan where CEOs care and the incentives are set so people are often doing the right thing).

ismailmaj··on Ask HN: What are you working on? (February 2026)
I agree that culture cascades but leadership is not the top of the food chain, their decisions or culture is often influenced by the capital.

The meta recently has been to ignore everything and only focus on maximizing the stock price, which is aligned with their incentives (stock based compensation, bonus based on stock performance) and the modern investors (money quick now).

I wonder if the rotten feeling I have on management is linked to western societies moving from high trust to low trust and from patient to impatient.

ismailmaj··on Ask HN: What are you working on? (February 2026)
How do you scale a more human management system when both the manager or the managee can be incompetent or malicious.

I'm aware that incentives are often counterproductive or insulting for white collar workers where there are elements of creativity in the job, but I feel like to get it to work you need trust which IMO cannot be scaled to big companies like Google.

FYI, I'm paraphrasing, perhaps badly "The puzzle of motivation | Dan Pink | TED". https://www.youtube.com/watch?v=rrkrvAUbU9Y

ismailmaj··on Are You YES AI or No AI?
Only after voting I understood it was related to duckduckgo....
ismailmaj··on Resistance training load does not determine hypertrophy
Even if true, high rep is impractical, otherwise we'd see people doing body weight exercises only reach high levels of bodybuilding.

Even around 1900, it didn't matter if you were a genetic freak, you needed a barbell to win competitions.

ismailmaj··on Rust--: Rust without the borrow checker
undefined behavior on steroids be like:
ismailmaj··on On privacy and control
Ironically, if your setup is too niche (e.g. browsing privacy configuration) you can be easily tracked, though no one will bother, but captcha's will certainly not miss you.
ismailmaj··on On privacy and control
My next low hanging fruit is certainly to make my LLM usage local, my queries contain much more sensitive information than what is mentioned by this post.

In the past I dropped off privacy when it was too inconvenient. For example I dropped protonmail because of bad search, left Linux desktop for Windows due to missing software, etc, I still haven't found the sweet spot for LLMs yet.

For the rest, I'm currently running the full macOS, iOS, safari, Apple passwords and I'm decently happy with this middle ground.

ismailmaj··on I asked AI researchers and economists about SWE career strategies given AI
fails on Safari macOS, works on chrome though.
← PreviousPage 3 of 5Next →