jerpint

30 ms·

391.
by jerpint 4y ago
We will be adding that soon
392.
by jerpint 4y ago
Run it using doc-ker
393.
by jerpint 4y ago
The hard part will be figuring out your different categories to classify. As some have pointed out, embedding your documents using an LLM would be a good bet. If you take the time to manually annotate a portion of your data, you could then
394.
by jerpint 4y ago
Having it be on par with DaVinci is a big claim, will definitely be putting that to the test
395.
by jerpint 4y ago
This is probably a few weeks of iteration away of being able to run on regular consumer hardware
396.
by jerpint 4y ago
You can checkout our library too which does just that :) https://github.com/jerpint/buster
397.
by jerpint 4y ago
Attention doesn’t have a window, unless you mean there’s a limit to the number of absolute positional embeddings available at train time?
398.
by jerpint 4y ago
That’s a great question - typically memory in transformers scale as O(N^2) with token count, so there must be an upper limit, but I would bet it to be far more than 4096 tokens
399.
by jerpint 4y ago
There have been CPU implementations of LLAMA (7b parameters, comparable in size) with very impressive performance
400.
by jerpint 4y ago
Without root it wouldn’t work
401.
by jerpint 4y ago
You could just prompt chatGPT more explicitly without giving it full access to your system
402.
by jerpint 4y ago
The DAN jailbreak relied on the concept of death to get chatGPT to comply
403.
by jerpint 4y ago
While this is very impressive results, they’re not actually having the models do any of the robotic controls. They only have the LLMs output text of what the controls should be, which then gets translated to actual movements. This works in
404.
by jerpint 4y ago
Anyone aware of a mobile app doing something similar? I guess I could use this too vis browser
405.
by jerpint 4y ago
While this is very impressive results, they’re not actually having the models do any of the robotic controls. They only have the LLMs output text of what the controls should be, which then gets translated to actual movements. This works in
406.
by jerpint 4y ago
Very suspect
407.
by jerpint 4y ago
The section that explains why python is good for AI talks about pybrain, a library that seems to date from 10 years ago. I’m pretty well versed in most ml frameworks and never heard of it. Last update to the website looks to be 2010. Weird
408.
by jerpint 4y ago
Im a heavy user of GitHub website and I find it doesn’t stop improving
409.
by jerpint 4y ago
Yes but in their wake they still left great contributions (CLIP, whisper, etc). So far they still seem net positive, even if they’re now for profit
410.
by jerpint 4y ago
Not to mention without needing expertise to deploy the thing
411.
by jerpint 4y ago
I meant more transformers than LLMs, models trained in self supervised way like DINO actually are pretty good at image segmentation
412.
by jerpint 4y ago
Multimodal transformers is the logical leap forward. These ideas are not new, deep mind and google have been at this for some time. It’s yet another arms race and those with the most diverse data will win
413.
by jerpint 4y ago
Transformers were used on text to train models without needing labeled data. People realized that simply scaling the data and models meant better performance. when they scaled even further, emergent intelligence started appearing and the mo
414.
by jerpint 4y ago
I’m a big fan of vimium! Works great and consistent
415.
by jerpint 4y ago
Big if. It’s been a few days they’ve released the model, and no one I know (academics) has gotten access to it yet
416.
by jerpint 4y ago
Except chatGPT has many obvious use cases
417.
by jerpint 4y ago
It’s PR on top of PR. OpenAI is “dedicated to making ai open TM” but also to making AI safe, so for “our safety”, opts not to release models. Not for business reasons /s
418.
by jerpint 4y ago
I just signed some form with license to maybe possibly get access to the model… not very open indeed
419.
by jerpint 4y ago
The summary itself is still generated, but has all the context to do summarization in the prompt. It's very difficult to otherwise finetune existing LLMs. GPT itself is closed-sourced, and doesn't allow for finetuning (except via
420.
by jerpint 4y ago
Nice! We built something very similar recently, it is more like "Ask your documentation" but very similar implementations otherwise. See a demo on the huggingface transformers documentation: https://huggingface.co/

More ›