Not this generation of AI though. It's a text predictor, not a logic engine - it can't find actual flaws in your code, it's just really good at saying things which sound plausible.
I can tell from this statement that you don't have experience with claude-code.
It might just be a "text predictor" but in the real world it can take a messy log file, and from that navigate and fix issues in source.
It can appear to reason about root causes and issues with sequencing and logic.
That might not be what is actually happening at a technical level, but it is indistinguishable from actual reasoning, and produces real world fixes.
It is true that models can happen to produce a sound reasoning process. This is probabilistic however (moreso than humans, anyway).
There is no known sampling method that can guarantee a deterministic result without significantly quashing the output space (excluding most correct solutions).
I believe we'll see a different landscape of benefits and drawbacks as diffusion language models begin to emerge, and as even more architectures are invented and practiced.
I have a tentative belief that diffusion language models may be easier to make deterministic without quashing nearly as much expressivity.
Citation needed.
I am sure the output of current frontier models is convincing enough to outperform the appearance of humans to some. There is still an ongoing outcry from when GPT-4o was discontinued from users who had built a romantic relationship with their access to it. However I am not convinced that language models have actually reached the reliability of human reasoning.
Even a dumb person can be consistent in their beliefs, and apply them consistently. Language models strictly cannot. You can prompt them to maintain consistency according to some instructions, but you never quite have any guarantee. You have far less of a guarantee than you could have instead with a human with those beliefs, or even a human with those instructions.
I don't have citations for the objective reliability of human reasoning. There are statistics about unreliability of human reasoning, and also statistics about unreliability of language models that far exceed them. But those are both subjective in many cases, and success or failure rates are actually no indication of reliability whatsoever anyway.
On top of that, every human is different, so it's difficult to make general statements. I only know from my work circles and friend circles that most of the people I keep around outperform language models in consistency and reliability. Of course that doesn't mean every human or even most humans meet that bar, but it does mean human-level reasoning includes them, which raises the bar that models would have to meet. (I can't quantify this, though.)
There is a saying about fully autonomous self driving vehicles that goes a little something like: they don't just have to outperform the worst drivers; they have to outperform the best drivers, for it to be worth it. Many fully autonomous crashes are because the autonomous system screwed up in a way that a human would not. An autonomous system typically lacks the creativity and ingenuity of a human driver.
Though they can already be more reliable in some situations, we're still far from a world where autonomous driving can take liability for collisions, and that's because they're not nearly as reliable or intelligent enough to entirely displace the need for human attention and intervention. I believe Waymo is the closest we've gotten and even they have remote safety operators.
I'm not sure if I'm up to date on the latest diffusion work, but I'm genuinely curious how you see them potentially making LLMs more deterministic? These models usually work by sampling too, and it seems like the transformer architecture is better suited to longer context problems than diffusion
I happen to use it on a daily basis. 4.6-opus-high to be specific.
The other day it surmised from (I assume) the contents of my clipboard that I want to do A, while I really wanted to B, it's just that A was a more typical use case. Or actually: hardly anyone ever does B, as it's a weird thing to do, but I needed to do it anyway.
> but it is indistinguishable from actual reasoning
I can distinguish it pretty well when it makes mistakes someone who actually read the code and understood it wouldn't make.
Mind you: it's great at presenting someone else's knowledge and it was trained on a vast library of it, but it clearly doesn't think itself.
The suggestion it gave me started with the contents of the clipboard and expanded to scenario A.
Being honest; I probably have to write some properly clever code or do some actual design as a dev lead like… 2% of my time? At most? The rest of the code related work I do, it’s outperforming me.
Now, maybe you’re somehow different to me, but I find it hard to believe that the majority of devs out there are balancing binary trees and coming up with shithot unique algorithms all day rather than mangling some formatting and dealing with improving db performance, picking the right pattern for some backend and so on style tasks day to day.
The idea that the entire top down processes of a business can be typed into an AI model and out comes a result is again, a specific type of tech person ideology that sees the idea of humanity as an unfortunate annoyance in the process of delivering a business. The rest of the world see's it the other way round.
Here's the most approachable paper that shows a real model (Claude 3 Sonnet) clearly having an internal representation of bugs in code: https://transformer-circuits.pub/2024/scaling-monosemanticit...
Read the entire section around this quote:
> Thus, we concluded that 1M/1013764 represents a broad variety of errors in code.
(Also the section after "We find three different safety-relevant code features: an unsafe code feature 1M/570621 which activates on security vulnerabilities, a code error feature 1M/1013764 which activates on bugs and exceptions")
This feature fires on actual bugs; it's not just a model pattern matching saying "what a bug hunter may say next".
PS: I know it is interesting and I don't doubt Antrophic, but for me it is so fascinating they get such a pass in science.
The lifeblood of the field is proof-of-concept pre-prints built on top of other proof-of-concept pre-prints.
We're in the "mad science" regime because the current speed of progress means adding rigor would sacrifice velocity. Preprints are the lifeblood of the field because preprints can be put out there earlier and start contributing earlier.
Anthropic, much as you hate them, has some of the best mechanistic interpretability researchers and AI wranglers across the entire industry. When they find things, they find things. Your "not scientifically rigorous" is just a flimsy excuse to dismiss the findings that make you deeply uncomfortable.
There's a ton of peer reviewed papers on SAEs in the past 2 years; some of them are presented at conferences.
For example: "Sparse Autoencoders Find Highly Interpretable Features in Language Models" https://proceedings.iclr.cc/paper_files/paper/2024/file/1fa1...
"Scaling and evaluating sparse autoencoders" https://iclr.cc/virtual/2025/poster/28040
"Identifying Functionally Important Features with End-to-End Sparse Dictionary Learning" https://proceedings.neurips.cc/paper_files/paper/2024/hash/c...
"Gemma Scope: Open Sparse Autoencoders Everywhere All At Once on Gemma 2" https://aclanthology.org/2024.blackboxnlp-1.19.pdf
You don't think a pattern matcher would fire on actual bugs?
On the flip side the idea of this being true has been a very successful indirect marketing campaign.
I don’t think we even have a coherent definition of human intelligence, let alone of non-human ones.
If you can't speak, can you think? Yes. Large Language model. Thinking is not predicated on language.
A few good starts for you. Please refute all of these arguments in your spare time to convince me otherwise:
* https://machinelearning.apple.com/research/illusion-of-think... * https://archive.is/FM4y8 * https://www.raspberrypi.org/blog/secondary-school-maths-show...
My point was that we don’t have a great definition for (human) intelligence either. The articles you posted also don’t seem to be too confident in what human intelligence actually entails.
See https://en.wikipedia.org/wiki/Intelligence
> There is controversy over how to define intelligence. Scholars describe its constituent abilities in various ways, and differ in the degree to which they conceive of intelligence as quantifiable.
Given that an LLM isn’t even human but essentially an alien entity, who can confidently say they are intelligent or not?
I’m very sceptic of those who are very convinced one or the other way.
Are LLMs intelligent in the way that humans are? I’m quite sure they aren’t.
Are LLMs just stochastic parrots? I don’t find that framing convincing anymore either.
Either way it’s not clear, just check how this topic is discussed daily in most frontpage threads for the last couple of years
Also, designers of these systems appear to agree: when it was shown that LLMs can't actually do calculations, tool calls were introduced.
The same goes for a lot of bugs in code. The best prediction is often the correct answer, being the highlighting of the error. Whether it can "actually find" the bugs—whatever that means—isn't really so important as whether or not it's correct.
Again - they're very useful, as they give great answers based on someone else's knowledge and vague questions on part of the user, but one has to remain vigilant and keep in mind this is just text presented to you to look as believable as possible. There's no real promise of correctness or, more importantly, critical thinking.
Intelligence, can be borne of simple targets, like next token predictor. Predicting the next token with the accuracy it takes to answer some of the questions these models can answer, requires complex "mental" models.
Dismissing it just because its algorithm is next token prediction instead of "strengthen whatever circuit lights up", is missing the forest for the trees.
To the nay sayers... good luck. No group of people's opinions matter at all. The market will decide.
A large quantity of bugs as raised are now fixed by claude automatically from just the reports as written. Everything is human reviewed and sometimes it fixes it in ways I don't approve, and it can be guided.
It has an astonishing capability to find and fix defects. So when I read "It can't find flaws", it just doesn't fit my experience.
I have to wonder if the disconnect is simply in the definition of what it means to find a flaw.
But I don't like to argue over semantics. I don't actually care if it is finding flaws by the sheer weight of language probability rather than logical reasoning, it's still finding flaws and fixing them better than anything I've seen before.
I feel that many people that don't find AI useful are doing things like, "Are there any bugs in this software?" rather than developing the appropriate harness to enable the AI to function effectively.
I think some of us come to terms with it in different ways.
Simpler, more mundane (not exactly, still incredibly complicated) stuff like homeostasis or motor control, for example.
Additionally, our ability to plan ahead and simulate future scenarios often relies on mechanisms such as memory consolidation, which are not part of the whole pattern recognition thing.
The brain is a complex, layered, multi-purpose structure that does a lot of things.