imranq

This just seems like a fancy way of describing LoRA? At the end of the day you are still learning weights based on a described set of outputs and then applying them to inference

There are already some resources like that like:

leetgpu.com

https://github.com/srush/GPU-Puzzles

For me its about a sense of progress, like in chess you can have an ELO score. Or in Duolingo theres a roadmap. If there were levels to this you could get more confident in your abilities.

Right now the levels are basically bachelors, masters, and PhD. Coarse and expensive

If someone could turn these into an adaptive Khan Academy style app, that would be incredible

I wish every paper had a succinct image like that!

Dang only forward passes. The real secret was in the backward pass! I was also curious to learn how they implemented the dualpipe scheduler

Amazing, this is why I visit HN :D

Combining the two ideas, "hand drawn" animation flip books could be neat. Start with a video, convert frames to a drawable vector, draw each frame with a pen plotter, assemble the flip book

Love it - I wonder if you could create a animation book this way cheaply, see a moving image as you flip through it

I think its amazing that posts like this exist, and more should definitely be written so that people don't feel powerless after a layoff. Too often we tie our identity to institutions and it isn't doing anyone any good (well maybe it helps the shareholders).

Based on the chart, the Kaggle SOTA model is far more impressive. These O3 models are more expensive to run than just hiring a mechanical turk worker. It's nice we are proving out the scaling hypothesis further, it's just grossly inelegant.

The Kaggle SOTA performs 2x as well as o1 high at a fraction of the cost

Summary: Its a real result, the cool part is more qubits seem to live longer rather than shorter, bad part the results are not explicitly verifiable, only through extrapolation

I think this is a great way to explain the idea of a knowledge graph. There's probably a whole field of ux research opening up to interact with knowledge graphs and using llms to create them, parse them and display them to users

So this is flipping the Human-AI working model and basically using the human as the tool?

This presentation does a good job distilling why FireDucks is so fast:

https://fireducks-dev.github.io/files/20241003_PyConZA.pdf

The main reasons are

* multithreading

* rewriting base pandas functions like dropna in c++

* in-built compiler to remove unused code

Pretty impressive especially given you import fireducks.pandas as pd instead of import pandas as pd, and you are good to go

However I think if you are using a pandas function that wasn't rewritten, you might not see the speedups

What happens when the enemy is deemed to be within?

Wow this is great! I think this is kind of the future of automation and "computer use" once LLMs become powerful enough.

Every task on the web can be reduced down to a series of backend calls, and the key is extracting out the minimal graph that can replicate that task.

Apple should hook into app functions themselves instead of relying on UI. I would be really surprised if Apple made a browser automation tool, since that would be the complete opposite of the "it just works" credo

I'd argue that this is not even a solution to begin with. If the LLM gets even one pixel value wrong, then at best the whole process breaks down. At worst, you could do some irreversible damage.

This is basically RPA with LLMs. And RPA is basically the worst possible solution to any problem.

Agents won't get anywhere because any user process you want to automate is better done by creating APIs and creating a proper guaranteed interface. Any automated "computer use" will always be a one-off, absurdly expensive, and completely impractical.

This is great -- I'm always amazed how effective classical algorithms are at doing so-called "neural tasks". What's strange is how few SIGGRAPH tech ever makes it out as a consumer product

Kind of lame to go to a holy site while basically spitting on the people who give you hospitality. "Unpleasant garbage" indeed.

Read Leopold Weiss' book (known as Muhammad Asad), "Road to Mecca", which is a much better memoir. A beautiful tale of self-discovery and bridging cultures and religion

I see where the author is coming from - like any new technology, AI is getting abused significantly: Better scams, stealing creator IP, developing thought crutches, emotional dependencies on artificial companions, vast energy usage, and the collapse of personality and knowledge across the societal mean.

Some people say that LLMs make you more productive, but guess what? The value captured by your productivity will only enrich the companies that employ you. Companies will hire fewer people and eliminate positions all for "shareholder value" and to give a small raise to a C-suite executive.

But at the same time, we could definitely envision a world where these models accelerate human creativity - smaller high performance LLMs working on the edge, responsible usage for learning and skill development, helping people build their own voice.

The key is to identify the subtle line between dependence and empowerment, and to know when one is accruing AI-debt at the expense of their own abilities.

The jets are 23 million light years in length! That's 140 milky way galaxies laid out -- these are sizes I can't even begin to comprehend

Nice. I think in the future this could be way better if everything was local and didn't require a API key. As far as I can tell mem0 is a fancy retrieval system. It could probably work pretty well locally with simpler models

Wow that's cool - that's like those adversarial attacks they do on self-driving cars to make it think a stop sign is a gorilla

How do you know what noise to add?

I watch all my lectures and practice language learning apps like Duolingo on the treadmill. It works!

This seems like fake news. It would extremely dumb to have such a policy since it would eventually be leaked and be negative press

I think HN is a community where people want to post something novel or new. When someone wants to post a kudos, most likely they'll upvote someone else instead of posting yet another "awesome job" (even if it is certainly warranted). Criticism instead can be endlessly diverse since there's usually only limited number ways to get it right, but plenty to get wrong.

In the end, HN comments fall prey to this truth and you see a handful of positive comments, with the majority being criticisms or "I wish this did X". No one person is to blame. Its just the culture of technologists today.

"My algorithm has always been: You put smart people together, you give them a lot of freedom, create an atmosphere where everyone talks to everyone else. They're not hiding in the corner with their own little thing. They talk to everybody else. And you provide the best infrastructure. The best computers and so on that people can work with and make everyone partners"

You could even go up a few iPhone versions and still be <$199. iPhone 12 used ~$190 supports iOS 17 and can run the Tiny models (phi-2).

For small models, today a used iphone 14 costs ~$300. Which means next year it will by down to ~$200.

I'm pretty sure by 2026, we will be able to inference "medium" models like Mistral/Llama3 8B on a <$199 device. That's not even factoring in how much better models will get by that time

Take a look at an app like this: https://apps.apple.com/us/app/offline-chat-private-ai/id6474...

This will perform inference on Phi-2 or Mistral on almost any iPhone device. The first neural engine came with the A11 bionic chip which is in the iPhone 8. People can buy that phone for $150 used on Amazon.

I think that's evidence that inference on the edge is already here