Learning with Shikhu, Model shakeups, and inferencing
Newsletter number #3, we’re getting there!
Welcome to the third newsletter blast for Answering Machines. Yes, I took even longer this time. Yes i’m still doing this, and we’re gonna get a bit more consistent. Ya boy has been cooking!
A new tool for myself and others
This week’s/months/quarters article is an intro to my new project Shikhu. Shikhu is a CLI tool and Agent Skill that helps you learn the codebases you write with agents. This is a continuation on my thoughts in the last blog post I sent about learning with agents, and a tool I hope helps you as it does me. I launched the tool about a week or so ago to a lot of support (thank you!) and I already have a few people using!!
Read more here: https://www.arjunkirtipatel.com/blog/introducing-shikhu
By the way, if you got this email from someone else, or found it on the web, please subscribe. I promise, it’ll be a good time.
Updates on Shikhu, if available. Changelog, or details
New in the newsletter is that I’ll be including a bit of a changelog in Shikhu too. Right now, there’s not much to report, although I’m planning to fix a bug related to multiple choice questions having A as an answer often, as well as exploring tools like Entire for developing features.
Next, some interesting bits of the internet I’ve collected for your reading pleasure. There’s been sooooo much stuff that’s happened since the last newsletter, so I tried to pick some things you probably haven’t picked up on here:
—-
Do.. do we really think it’s gonna stop with Mythos?
Arjun’s Take: (above is a linkedin post I thought was a really solid summary of how I feel)
There has been a lot of odd, almost hypocritical noise made about the Mythos ban. Particularly, it seems a lot of folks have a “just-desserts” style view of what’s happened to Anthropic, and suggesting a change to open source models instead. I think open souce models are awesome, we need more of them, but I think this is not a good takeaway.
I don’t think the ban is a good idea, and I don’t think it comes from a place where good ideas live. I also don’t see why it wouldn’t extend to the frontier-level open source models, especially those that aren’t US-made. So I think we should have clever-er regulation rather than, say, a personalist view of what should/shouldn’t go….
Unconventional AI develops a network of physical-style oscillators to generate images
Turns out, you can pick physical analogues of systems to “train” networks instead of just weights!
Arjun’s Take: Truely wild innovation. Like, literally, they made “virtual pendulums” that are all attached to one another, who’s movement correlates to the progagation that occurs in typical neural networks. Somehow is this way more energy efficient than typical networks and works to generate images. It’ll probably be a while until something like this becomes competitive with production models now, but it’s so cool that they basically figured out the math for a physical model of learning. Highly recommend the writeup as it as some sick visuals!
Code2Lora: Relearning codebases using adapters instead of prompting
A researcher’s linkedin post on the project, in addition to codebases/demos
Arjun’s Take: I hate reprompting my codebases for my LLMs, so this is a neat take. Need to explore it to further understand, but worth showing here!
Sakana AI and OpenRouter release “fusion” endpoints
Both companies announced slightly different takes to combine LLMs to work better together
Arjun’s Take: We’re gonna see more and more techniques like these as people get sensitive to token costs and want to diversify from adversarial US action. I don’t know which approach is better, but worth exploring both!
Inference providers Baseten and Modal rise, with hosted open source model endpoints
Modal launches Auto Endpoints, and Baseten raises a huge funding round to encourage diversification
Arjun’s Take: Speaking of open source models, I think Baseten/Modal have the right approach here. They are inference providers that have optimized, custom deployments of open source models on the cheap. So you get a standardized endpoint, serving expertise, and freedom to upgrade. I wanna try this for sure. But, I’m not sure it’s great to separate the model training companies from the inferencing ones, so unless prices never drop from Anthropic/OpenAI, I’m curious to see what innovation these companies will introduce.
A quick note for this week. I’m really proud of myself, finally releasing Shikhu as a project. I really think it has some legs, but I need more feedback from others to understand what to do with it. If learning your codebases sounds like something you think is worth doing, I ask you take a look, or at least shoot me a star for my troubles. Thanks all!
