Aider actually prompts the model to say if it needs to see additional files. Whenever the model mentions file names, aider asks the user if they should be added to context.
As well, any files or symbols mentioned by the model are noted. They influence the repomap ranking algorithm, so subsequent requests have even more relevant repository context.
This is designed as a sort of implicit search and ranking flow. The blog article doesn’t get into any of this detail, but much of this has been around and working well since 2023.
I see, so the context adapts as the LLM interacts with the codebase across requests?
That's a clever implicit flow for ranking.
The difference in my approach is that exploration is happening within a single task, autonomously. The agent traces through structure, symbols, implementations, callers in many sequential lookups without human interaction. New files are automatically picked up with filesystem watching, but the core value is that the LLM can navigate the code base the same way that I might.
Are you using LLM to help you write these replies, or are you just picking up their stylistic phrasings the way expressions go viral at an office till everyone is saying them?
As an LLM, you wouldn't consider that you're replying confidently and dismissively while clearly having no personal experience with the CLI coding agent that not only started it all but for a year (eternity in this space) was so far ahead of upstarts (especially the VSCode forks family) it was like a secret weapon. And still is in many ways thanks to its long lead and being the carefully curated labor of a thoughtful mind.
As a dev seeking to improve on SOTA, having no awareness of the progenitor and the techniques one most do better than, seems like a blind spot worth digging into before dismissing. Aider's benchmarks on practical applicability of model advancements vs. regressions in code editing observably drove both OpenAI and Anthropic to pay closer attention and improve SOTA for everyone.
Aider was onto something, and you are onto something, pushing forward the 'semantic' understanding. It's worth absorbing everything Paul documented and blogged, and spending some time in Aider to enrich a feel of what Claude Code chose to do the same or differently, which ideas may be better, and what could be done next to go further.
I’m building a quantum photonics experiment that is a variation of the quantum eraser.
One aspect that HN may find interesting is my use of Bayesian optimization to control and perfect key experimental settings. About a dozen of the wave plates and other optical components are motorized and under computer control.
Given a goal metric like "maximally entangle the photon pairs" the optimizer will run the experiment 50-100 times, tweaking the angles of various optics and collecting data. Ultimately it will learn to maximize the given cost function.
This sort of thing is commonly done with tools like Optuna during NN/LLM training to optimize hyper-parameters, but seems less common in physics especially quantum photonics. I'm using a great tool called M-loop to drive the optimization, which was originally developed for creating Bose-Einstein condensates.
I like the unconventional approach. A few minutes with GPT raises two issues:
1. We've raised CO2 from 280ppm to 420ppm, about a 50% increase. To dilute it back down would require 50% more total atmosphere. This would also raise the surface air pressure 1.5x.
2. How much heat is trapped is related to the absolute amount of CO2 in the atmosphere, not the fraction. So the diluted atmosphere would retain just as much heat.
It would be interesting to see how hard it would be to walk these models towards general relativity and quantum mechanics.
Einstein’s paper “On the Electrodynamics of Moving Bodies” with special relativity was published in 1905. His work on general relativity was published 10 years later in 1915. The earliest knowledge cuttoff of these models is 1913, in between the relativity papers.
The knowledge cutoffs are also right in the middle of the early days of quantum mechanics, as various idiosyncratic experimental results were being rolled up into a coherent theory.
> It would be interesting to see how hard it would be to walk these models towards general relativity and quantum mechanics.
Definitely. Even more interesting could be seeing them fall into the same trappings of quackery, and come up with things like over the counter lobotomies and colloidal silver.
On a totally different note, this could be very valuable for writing period accurate books and screenplays, games, etc ...
And it's a 4B model. I worry that nontechnical users will dramatically overestimate its accuracy and underestimate hallucinations, which makes me wonder how it could really be useful for academic research.
I think not everyone in this thread understands that. Someone wrote "It's a time machine", followed up by "Imagine having a conversation with Aristotle."
There's quite a lot of text in pre-Internet daily newspapers, of which there were once thousands worldwide.
When you're looking at e.g. the 19th century, a huge number are preserved somewhere in some library, but the vast majority don't seem to be digitized yet, given the tremendous amount of work.
Given how much higher-quality newspaper content tends to be compared to the average internet forum thread, there actually might be quite a decent amount of text. Obviously still nothing compared to the internet, but still vastly larger than just from published books. After all, print newspapers were essentially the internet of their day. Oh, and don't forget pamphlets in the 18th century.
> the issue is there is very little text before the internet,
Hm there is a lot of text from before the internet, but most of it is not on internet. There is a weird gap in some circles because of that, people are rediscovering work from pre 1980s researchers that only exist in books that have never been re-edited and that virtually no one knows about.
There is no doubt trillions of tokens of general communication in all kinds of languages tucked away in national archives and private collections.
The National Archives of Spain alone have 350 million pages of documents going back to the 15th century, ranging from correspondence to testimony to charts and maps, but only 10% of it is digitized and a much smaller fraction is transcribed. Hopefully with how good LLMs are getting they can accelerate the transcription process and open up all of our historical documents as a huge historical LLM dataset.
I’ve relied heavily on seamless capture for a couple of decades, ala Getting Things Done.
My solution is a twilio text number that automatically inserts any texts it receives into the top of my todo.md file. Previously todo.org, until about a year ago.
iOS has ubiquitous support to quickly share to SMS from any/everywhere. It’s easy to send a text to this contact from a Home Screen shortcut, but also also from the share sheet in most every app.
I’ve been building proficiency with quantum optics equipment. Repeating classic quantum entanglement experiments like the quantum eraser [0] and violating the CHSH inequality (which won the 2022 Nobel). I’m working towards a novel quantum eraser variant.
I really like LLM+sympy for math. I have the LLM write me a sympy program, so I can trust that the symbolic manipulation is done correctly.
The code is also a useful artifact that can be iteratively edited and improved by both the human and LLM, with git history, etc. Running and passing tests/assertions helps to build and maintain confidence that the math remains correct.
I use helper functions to easily render from the sympy code to latex, etc.
A lot of the math behind this quantum eraser experiment was done this way.
Below are some great videos on the physics and practicalities of single mode fiber. They are Thorlabs videos, so are slanted more towards the use of SMF in a laser lab rather than a telecom setting. They reference a lot of the theory, but also provide a good intuition about how and why SMF works so well.
As well, any files or symbols mentioned by the model are noted. They influence the repomap ranking algorithm, so subsequent requests have even more relevant repository context.
This is designed as a sort of implicit search and ranking flow. The blog article doesn’t get into any of this detail, but much of this has been around and working well since 2023.
reply