The Second Half
ysymyth.github.io
ysymyth.github.io
I used to think the problem was that models can't learn over time like humans, but maybe that can be worked around. Today's models have large enough context windows to fit a medium sized project's complete code and documentation, and tomorrow's may be larger; good-enough world knowledge can be maintained by re-training every few months. The real problem is that even models with large context windows struggle with complexity moreso than humans; they miss crucial details, then become very confused when trying to correct their mistakes and/or miss other crucial details (whereas humans sometimes miss crucial details, but are usually able to spot them and fix them without breaking something else).
Reliability is another issue, but I think it's related to scalability: an LLM that cannot make reliable inferences from a small input data, cannot grow that into a larger output data without introducing cascading hallucinations.
EDIT: creative control is also superseded by reliability and scalability. You can generate any image imaginable with a reliable diffusion model, by first generating something vague, then repeatedly refining it (specifying which details to change and which to keep), each refinement closer to what you're imagining. Except even GPT-4o isn't nearly reliable enough for this technique, because while it can handle a couple refinements, it too starts losing details (changing unrelated things).
With English, the meaning of a sentence is mostly self-contained. The words have inherent meaning, and if they’re not enough on their own, usually the surrounding sentences give enough context to infer the meaning.
Usually you don’t have to go looking back 4 chapters or look in another book to figure out the implications of the words you’re reading. When you DO need to do that (maybe reading a research paper for instance), the connected knowledge is all at the same level of abstraction.
But with code, despite it being very explicit at the token level, the “meaning” is all over the map, and depends a lot on the unwritten mental models the person was envisioning when they wrote it. Function names might be incorrect in subtle or not-so-subtle ways, and side effects and order of execution in one area could affect something in a whole other part of the system (not to mention across the network, but that seems like a separate case to worry about). There’s implicit assumptions about timing and such. I don’t know how we’d represent all this other than having extensive and accurate comments everywhere, or maybe some kind of execution graph, but it seems like an important challenge to tackle if we want LLMs to get better at reasoning about larger code bases.
You can have a book where in the last chapter you have a phrase "She was not his kid."
Knowing nothing else, you can only infer the self-contained details. But in the book context this could be the phrase which turns everything upside down, and it could refer to a lot of context.
It makes me think about things like... "what if we also provided not just the source code, but the abstract syntax tree or dependency graph", or at least the related nodes relevant to what code the LLM wants to change. In this way, you potentially have the true "full" context of the code, across all files / packages / whatever.
As a concrete example, a random bit of code from the minih264 encoder:
/**
* Quantized/dequantized representation for 4x4 block
*/
typedef struct
{
int16_t qv[16]; // quantized coefficient
int16_t dq[16]; // dequantized
} quant_t;
LLMs already have a lot of this knowledge, and could probably answer if prompted, but my point is more that the code doesn't explicitly lay out all of these things unless you carefully trace the execution, and even then, some of the requirements might not be evident. Maybe negative numbers aren't valid here (I don't actually know) but the reason that invariant gets upheld is an abs() call 6 levels up the call stack, or the data read from the file is always positive so we just don't have to worry about it. I dunno.
Anyway I imagine LLMs could be even more useful if they knew more about all this implicit context somehow, and I think this is the kind of stuff that just piles up as a codebase gets larger.
- ReAct
- Reflexion
- SWE-bench
- OpenAI Deep Research
- OpenAI Operator
> I call this the utility problem, and deem it the most important problem for AI.
> Perhaps we will solve the utility problem pretty soon, perhaps not. Either way, the root cause of this problem might be deceptively simple: our evaluation setups are different from real-world setups in many basic ways.
LLMs are reaching the same stage that most exciting technologies reach. They have quickly attracted lots of investor money, but that is going to have to start turning into actual money. Many research papers are being written, but people are going to start wanting to see actual improvements, not just theoretical improvements on benchmarks.
For instance I ask AI assistants a lot about what some code is trying to do in applications software where it is a matter of React, CSS and how APIs get used. Frequently this is a matter of pattern matching and doesn't require deep thought and I find LLMs often nail it.
When it comes to "what does some systems oriented code do" now you are looking at halting problem kind of problems or cases where a person will be hypnotized by an almost-bubble-sort to think it's a bubble sort and the LLM is too. You can certainly make code understanding benchmarks aimed at "whiteboard interview" kind of code that are arbitrarily complex, but that doesn't reflect the ability or inability to deal with "what is up with this API?"
Which is, of course, true. Anecdotally, a lot of value I get from Copilot is in simple, mundane tasks.
Even in scientific domains where a high level of background knowledge is expected, it is standard practice to define each acronym prior to its use in the rest of the paper, for example “using three-letter acronyms (TLAs) without first defining them is a hindrance to readability.”
In the usage in the book/page, it seems to refer to how tasks/problems are run in parallel and the learning averaged, whereas the author is advocating these problems run sequentially.
https://en.wikipedia.org/wiki/Independent_and_identically_di...
The players it played against had never played against something that behaved so weirdly. It had lightning reflexes and it clearly wasn't human. It was playing a toy game mode requiring about 5% of skills needed for a full match. I'm other words, they engineered it to look good at they toy task, and it did. But they didn't give the pros any time at all to learn their opponent -- after all they might have figured out how to play against it!
Which is great! There's room in the world for new benchmarks that test for more diverse things!
It's highly likely at least one of the new benchmarks will eventually test for all the criteria being mentioned.
To say we are at a point where AI can do anything reliably is laughable, it can do much and it will tell you any answer whether right or wrong with full confidence. To trust such a technology in the big no-human decisions like we want it to is foolswork.
I hate acronyms with a fierce passion.
Please, let it be Rocket Launcher for once.