152334H

Regardless of what you want to learn, "small daily activities" is a bit hard. You can learn some stuff by osmosis, following the feeds of AI devs && AI channels, but the bulk of what I learn comes from starting projects & digging into code & reading papers.

If you can hold attention span over several days (I can't), work on a project bit-by-bit. Just make sure it uses modern AI stuff, and that you have smart people to talk around with.

I don't have much experience myself. I only started ~10 months ago -- just a month or two before Stable Diffusion.

You just have to do it every day. It's fun!

It'll work! I just haven't touched any of the 4bit stuff myself, so I don't personally know how to add it. Great low-hanging fruit for anyone else to take on.

It's better because

1. it's using vicuna as a base.

2. It has a pretty high quality fine-tuning dataset. I initially missed this, and it's a very important advantage.

3. (speculatively) it doesn't collapse to extremely short responses (which BLIP2 and other models trained on image-text caption pairs) because of how small/simple the adapter is.

I was interested in training a BLIP2-LLaMA model before this, and I might still do it just to test (3).

On a technical level, they're doing something really simple -- take BLIP2's ViT-L+Q-former, connect it to Vicuna-13B with a linear layer, and train just the tiny layer on some datasets of image-text pairs.

But the results are pretty amazing. It completely knocks Openflamingo && even the original blip2 models out of the park. And best of all, it arrived before OpenAI's GPT-4 Image Modality did. Real win for Open Source AI.

The repo's default inference code is kind of bad -- vicuna is loaded in fp16 so it can't fit on any consumer hardware. I created a PR on the repo to load it with int8, so hopefully by tomorrow it'll be runnable by 3090/4090 users.

I also developed a toy discord bot (https://github.com/152334H/MiniGPT-4-discord-bot) to show the model to some people, but inference is very slow so I doubt I'll be hosting it publicly.

Majority of deployed services right now are TorToiSe derivatives. https://git.ecker.tech/mrq/ai-voice-cloning

just fyi, this will change soon

Where A,B,C are empirical constants,

Loss = A/parameters + B/dataset_size + C

Just read the chinchilla paper.

I spoke with the authors of the paper; the leftmost points in Figure 1 were generated with batch-size 1, indicating ~1.2x and ~2x improvements in speed over DeepSpeed for 30B and 175B models respectively. For reference, this is speeding up from ~0.009tokens/s to about ~0.02tokens/s on 175B.

These results are generally unimpressive, of course. Most of the improvements at that point are attributable to the authors making use of a stripped down library for autoregressive sampling. HN falling for garbage once again...

HN is just incredibly bad at figuring out what kind of ML projects are worth getting excited about and what aren't.

MJ v4 doesn't even use Stable Diffusion as a base [0]; a fine-tune of the latter will never come close to achieving what they do.

[0] - https://discord.com/channels/729741769192767510/730095596861...

With all of these outages, you have to wonder if it's a lack of skilled engineers on their part, or if they simply don't have enough GPUs to keep the lights on all the time.

No mention of any other competitors that've been doing this stuff for several years? Uberduck? Fakeyou? Coqui? 15?

Another example of "ChatGPT for open-ended law": https://broodingomnipresence.substack.com/p/brooding-omnipre...

Tangentially related: DoNotPay's (probably illegal) plans to use LLMs in actual court sessions https://twitter.com/jbrowder1/status/1612312707398795264

More samples available on their github page:

https://valle-demo.github.io/

Personally, I find that their samples aren't near anything I'd call "dangerous". I cross-compared the baseline examples to the VALL-E ones when the paper dropped, and found several that were garbled in the usual robotics sounding TTS failures.

Probably a good thing that people are getting alarmed before a true indistinguishable voice cloner exists, though.

Forgive the anthropomorphising, but I find it very similar to how I'd handle someone forcing me to explain two contradictory claims coherently.

I see no one has mentioned Byte-Pair Encoding (BPE) yet.

https://www.gwern.net/GPT-3#bpes

The reason ChatGPT struggles with tasks involving individual character counting is that it can't see individual characters. The word "prime" is a single token (https://beta.openai.com/tokenizer) to the eyes of ChatGPT, and 1 is not prime.

Interesting, thanks.

Any word on what the neovim plugin does? Considering some features like,

> Then, most recently accessed 20 files of the same language are queried from VSCode.

are probably not available by default in nvim?

---

Also, do you think there's any chance you get in trouble for reverse engineering copilot?

There would be a lot less mundane questions on the internet if people would google/chatGPT first. I think it is valuable to try to direct people to using said tools when they are sufficient for an answer.

"Whisper V2" is a bit of an exaggeration. It's

* only an update to the `large` model (and therefore irrelevant for hobbyist, low CPU use cases)

* architecturally identical to the original whisper model

* just trained on more epochs

Not exactly a novel idea.

https://twitter.com/AdeptAILabs/status/1590396065072951296

I'm interested in finding out which day it'll start to break. GPT-3/ChatGPT are not perfectly accurate programmers, and some of the longer problems with more fluff text might distract it from the problem.

I would be really surprised if it solved up to day 10 with no mistakes.