ismailmaj

I think I would be horrified looking at your Claude API bill.

Apparently plus users do not have access to Sol, so I'm really worried about the ugly Terra Pelican.

I don't know what the day to day would be like but it was a ML research engineer role.

I'm pretty confident I would've been able to come up to the solution in better circumstances, maybe even without hints.

But it was clear in this case that the interviewer just took a question from the company's bank of questions and was alt-tabbed for half the interview, I have felt the energy early and I was also half-checked out.

I'm aware I'm saying this post factum, but I had a very fun first interview with that company and matched well with the first interviewer so my expectations were high, and then I got hit by the big tech style interview when it was an early stage startup.

I got that recently, I really didn't like that question.

For the n-th percentile version, the obvious solution is sorting and it takes 10 seconds to get to that point, 5 minutes of implementation with tests. Good. It's all downhill from here.

Then you get hit with the "it's a data stream" and you realize you have to implement a balanced tree on the spot which I wouldn't describe as fun.

You may or may not be able to implement that. I did not. Blabbered something about Rust having sorted B-Trees and I don't think Python has them -- they do not on the standard library.

Then the interviewer leaned heavily on the "reduce memory usage" and I couldn't come up with a solution (no shit it's Ω(n) and he didn't even tell me to go fetch for a randomized algorithm). I later understood he expected the reservoir sampling solution which is basically keeping a representative group of size K that is a good proxy of the whole stream, it goes like this: keep the K first elements, any elements after that replaces any element of our sample at random.

What I did after 10 minutes of weird silence is to assume the data stream follows a normal distribution and computing the P-percentiles by computing the running mean and standard deviation.

I felt frustrated at the end of the interview because it really felt like a big gotcha of either you know the reservoir sampling "leetcode trivia" or you don't.

High interest rates and instability makes it a bad environment to invest, it raises the bar on speculative/growth investment, and tech is mainly that.

Mix that with heavy AI bills, there isn't a lot of budget left for hiring.

From talking to someone that worked at YouTube for 15 years, they still had a lot of core Python code in 2016 that was legacy from the OG company/team and that code needed to be transitioned to follow the Google way of doing things in C++/Go.

I don't think it was distinct enough from the Google culture like Android was at the start of the acquisition but it seems they had leeway to do their own thing.

I'll pick small company, thank you.

I'm not getting the Opus 4.8 switch for coding, supposedly given how fast I reached the usage limit, which is kind of nice.

It's great for defining fancy floats used in machine learning

e.g. https://github.com/zml/zml/blob/33ced8fa078b3c7c8c709bd526ae...

they're still kings for training, though I've heard Anthropic is training now on JAX+TPU setup, so might not be a monopoly in that segment.

I can talk about France specifically, the working age population feels squeezed currently because a big portion of their wages goes to pay pensions, let's be specific:

for a 100k paid by your employer, 46k goes to you after all taxes, 45k goes to social contribution, of those 20k specifically to paying current pensioners[1].

It doesn't look so bad, unless you take into account the inversion of the demographic pyramid, the workers will get progressively more squeezed as older folks enter retirement, and are unlikely to get themselves good retirement given that there isn't enough kids to support them when they will be old.

And so, the rational move is to save more to fund your retirement while already getting pressured by relatively low wages and low saving potential due to the taxes, and France doesn't really have a growth lever to pull to get us out of this situation as we're already at a very high debt-to-GDP ratio.

So that's why some workers are unhappy with what they call "fat pensions", what also doesn't help in France is that the country is very centralized around Paris and rent is now another channel where money flows from workers to (often) pensioners.

As 2027 elections approaches, it seems we have the options of being even more fiscally irresponsible (the extremes), and the center seems to lean for austerity measures (Edouard Philippe) and none of these are making me optimistic about the future.

[1] https://code.travail.gouv.fr/outils/simulateur-embauche

The risk of being cut off from strategically important technologies, once minimal, is now very real.

Good timing given the Fable 5 story.

My main read on this is that European governments will more aggressively invest in regional AI labs.

For Nvidia chips you could've deluded yourself that the US is just anti-china, that position is harder to argue for right now.

In my experience a lot of companies try really hard to be data oriented and try to find objective metrics for impact, sometimes it’s good, often it’s bad. Like LOC count, PR count, time in meetings or time spent at the office.

Enough of this and people will learn to play the game over doing the right thing.

I had a much better experience gaming on MacOS than I ever did using Linux + X (it was first dual boot with Windows, then double GPU + VM, then proton).

But it's mainly because I do not target anymore games with stringent anti-cheats or with high setup requirements (though the M5 is quite powerful right now), and a lot of games are ported natively to macOS but not Linux (most recent to date is Age of Empire 2 Definitive Edition 1 week ago).

Gemini is a huge team while Gemma is relatively small. They can totally do this at a loss with no ulterior motive.

They remind me a bit of HuggingFace, create something great then make money … maybe.

I expected a 4090, maybe 2. I did not expect 8xH200 for a 2B model.

I just asked the model details about the incoming spaceX IPO and it responded with “There’s no confirmed SpaceX IPO. Elon Musk has said for years that SpaceX itself won’t go public”. It took me two push backs and specifically asking for web search.

I feel like I won’t like this model just like I didn’t like 4.7, push backs a lot and avoids thinking or search as much as possible.

3 things, they can, there is a precedent for that with Google v. Oracle for Java, and they have something!

AMD engineered something called HIP which is CUDA API compatible libraries that targets AMD's hardware, it's the closest thing we have for drop-in replacement to Nvidia's software moat.

It works for simple stuff but loses terribly for frontier kernels (like Flash Attention 3), novel approaches (e.g. Mamba) or networking (e.g. NCCL), also they are rough on the edges, so what you gain from GPU costs is lost in engineering cost.

My previous company tried to compete in this GPU game while putting effort to have a good software stack (Rivos), drop in replacement and cheaper with decent software.

But that vision was rough, any new player had to implement the bad APIs due to backward compatibility concerns, following specs wasn't sufficient as a lot of the AI stack was depending on observable effects (Hyrum's Law), and Nvidia simply just had a long head start, the company is now dead (acquired by Meta) and AFAIK there isn't another player.

Best case scenario AMD puts more effort into their software stack but I just think they do not have enough internal talent to compete.

Training will continue to be an Nvidia's thing and that's where most of the money sits, unless suddenly the AI research scene pivots to using JAX but I do not see it coming any time soon, if anything, I've seen internal efforts at Google to make PyTorch work nicely with TPUs. Some players like Anthropic started using JAX for training but all the small players are using Nvidia, I'm guessing it has something to do with Nvidia partnering aggressively with startups.

Their moat is cuda and cuda libraries and everything built on top.

When a new architecture drops, it's always PyTorch running on CUDA, other PyTorch backends are best effort, even if they reach feature parity, many industry power users went closer to the metal to squeeze performance and that stuff is too specific to Nvidia stuff.

if there is something that will beat Nvidia, it won't be something reaching feature parity with slightly better economics (like AMD, also Nvidia could just reduce their margins), it needs to be a novel approach worth rewriting the codebase for (maybe Cerebras, maybe a new player).

I'm thinking that personally, technology is not bad in a vacuum and not necessarily bad in society, but it just reveals that our system is ill-equipped to guarantee good usage of it.

We could have fun defining what's good usage but we're so far from it, it would just make me sad.

Unclear if it's the only cause but wafer scale is great for very low latency, but loses to throughput per dollar compared to classic Nvidia like GPUs. I don't think they can reduce the gap, SRAM is just more expensive than HBM and their architecture needs a lot of it.

So, the price makes it necessarily niche to some specific use-cases like HFT or intelligent duplex voice assistants, I'm still semi-bullish personally.

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.

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.

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.

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?

Oh no, cortisol spike in my text-only forum.

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).

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.