[hodgehog11]

In terms of price-to-productivity, nothing will beat DeepSeek right now. For what you have described, all of the existing frontier models will perform well (probably about the same even).

If you expanded the list to very hard research tasks, Fable was so far ahead of the others that it doesn't even deserve debate. If you are a researcher doing something involving scientific computation or mathematics that wasn't rejected by the guard rails, and you were using Fable, that week was probably your most productive week ever. A couple of my PhD students effectively finished their current projects in that period by getting Fable to chew on it for 30 hours straight (not sure how I feel about that).

[josephg]

This was my experience too. I asked fable to implement a (quite complex, novel) CRDT engine. It did fantastic work for the 3 days I had access to it. The spec it wrote is exactly what I want. It used prototypes and examples to figure out some hard problems and answer a lot of complex design questions. Claude Opus has been lumbering along over the last week trying to turn what it did into a useful library. As far as I can tell, by trying to vibe together pieces of work fable did without understanding it properly. Opus makes mistakes constantly, misunderstands the spec, and it makes terrible engineering decisions. Earlier today it claimed it proved something was impossible. I asked it to think that through and it immediately backtracked and apologized. Was it right then, or is it right now? Claude has no clue.

I'm kicking myself for not using Fable more while I had it. Now that I've used fable, I'm not sure I even want opus any more. It might be more efficient in the medium term to just program everything myself until I have access to a similarly capable model.

I feel like Deepseek, opus, etc are only good at problems that have already been solved 100 times on github. They're like the iPhone 3G. Its exciting they exist at all. But subsequent versions make them seem like cheap junk.

[manuisin]

I had a similar experience. For complex maths in my 3d engine around IK, made more progress in those 3 days than I would’ve done in weeks.

I recently discovered though that ChatGPT 5.5 Pro is almost as smart but only available to me via the ChatGPT app. I’ve been having it read my code and having Opus 4.8 use the ChatGPT app to collaborate. It’s a step down but its a temporary stopgap for complex problems.

[hodgehog11]

It's been pretty cool (and a bit scary) to hear similar experiences from many other researchers and developers just how impressive Fable was for their particular workflow and the productivity it seemed to unlock. Personally, I have found that I can't let any of the other AI agents touch the code that Fable produced for me, since they consistently fail to understand the very delicate choices that it made to ensure optimal performance.

[limflick]

> they consistently fail to understand the very delicate choices that it made to ensure optimal performance.

Do you have an example?

[gambiting]

I'm an engine programmer in video games on a big project - recently I'm mostly just fixing bugs. And Fable has been absolutely phenomenal in those 2 days we had it lol. It was genuinely the first time I didn't really have to reproduce a bug to fix it - Fable was able to really understand the code and interconnections between the systems to reason why it might have happened and fix it. Opus really struggles with that in my experience, or it munches code for 2 hours only to come up with a completely wrong explanation, like, wrong if you think about it for more than 2 seconds. TBF I only had 2 days to play with fable, but it was incredible in that time - can't wait to have it back.

[apatheticonion]

Something I wonder about in this domain is if this is the fault of the language.

Game development is usually in C++ or C#. With C++, bugs are a nightmare to find. C# less so, but there are still memory leaks possible.

I've been writing Rust for a few years now and it has been absolutely phenomenal. I get the performance of C++ and the only thing I think about when developing is the logic itself.

There aren't any mature game engines that use Rust (Bevy is rapidly improving, but it's no Unreal/Unity) - though you can compile Rust to a dll so it can be consumed by Unreal/Unity.

Due to the language constraints of Rust, I have found that LLMs need to work _way_ less to figure out bugs and render out code. The compiler gives very specific error messages and if it compiles, it works.

[gambiting]

So....just from my personal experience.

When we started getting access to Claude last year, it was barely competent at C++. Like someone who read a book about C++ but never actually wrote any, or had to debug any issue. It would confidently declare things which any C++ programmer could tell you were bollocks straight away.

But....now it's able to reason about things like memory management, thread safety and second order effects when multithreading very very very well. It's honestly incredible how well it can understand C++ code and figure things out. Opus is already pretty good(although far from perfect) but Fable was just something else. I could give it a description like "It was reported that only sometimes(1/20 repro) when the user opens the inventory and closes it, they can't move the camera afterwards" and it was able to find out both the inventory camera control and input code and pinpoint the exact race condition that could lead to this situation.

[throwaway2037]

Do you need to push internally to get access to these LLMs for writing code, or is mgmt aggressive and supportive?

[gambiting]

We have unlimited access to Claude and its use is very much encouraged across the entire organisation.

[steelframe]

Makes me wonder what the value of a PhD is.

[hodgehog11]

To train people to become researchers. If the definition of what a researcher is changes over time (due to AI or otherwise), then the nature of the PhD will also change over time.

[vincnetas]

I would say to push frontier of human knowledge forwards. Its not a high school anymore where you learn to learn. PhD is already a place where you should deliver value. (By "value" i don't mean commercial product)

[dataflow]

>> Makes me wonder what the value of a PhD is.

> I would say to push frontier of human knowledge forwards.

Not sure if this is what you meant, but to be clear: pushing human knowledge forward is a practical prerequisite of a PhD (given it's pretty hard to convince people you're adequately trained to do such a thing without actually demonstrating it), but it's not the value of the PhD itself. The value is producing a researcher -- that is, someone who has the skills to continue accomplishing this in the future.

The difference is that if you happen to expand human knowledge by a stroke of dumb luck (or smart AI...) without actually having acquired skills to continue doing so in the future, then you're not really earning a PhD.

[nonameiguess]

Depends on the field. There aren't really all that many research tasks that can be completed in a week. One of my favorite stories from the past few decades was the finding of accelerating expansion of space. It took three decades of scheduling telescope time, looking at exactly the right places in space at exactly the right time, to find enough supernovas to have data to even analyze. Not a whole lot of non-trivial science is just "read existing literature, think hard, and produce text."

[disgruntledphd2]

But honestly, there's a lot that is (particularly for simulation like stuff which can be implemented in code). We should expect to see a bunch of extra output here, which might improve scientific productivity (even though the real bottlenecks are gonna be the academic publishers and their fixation on holding on to their copyrights).

[data_maan]

Makes me wonder what the value of a human on a PhD course is

[Cakez0r]

Mimo 2.5 pro is the best intelligence / dollar

[atty]

In most cases it seems (at least to me and colleagues) to be turning out that picking best intelligence is a better option than picking better intelligence / dollar, assuming you can afford the cost. At least on interesting problems. If you’re doing generic web dev work, probably not the case.

[Bnjoroge]

What are interesting problems? Deepseek and some other open source models do decently well in the gpu kernel benches. https://kernelbench.com/hard Most people think they’re working on “very hard” problems and they really aren’t

[hodgehog11]

Absolutely agree with this. As Louis Rossmann recently pointed out, it's the difference between a correct answer and a wrong answer; the correct answer is worth a good amount, while the wrong answer is worth nothing. Under this metric, for harder tasks, the most intelligent model is best per dollar.

[apatheticonion]

That's why experience + cost effective model is IMO the best combination.

Experience allows you to design the skeleton where the implementation details are often inconsequential. There are relatively few scenarios where an LLM would need more guidance to render an outcome, but even dumb local models can do that.

Building a simple UI component vs an efficient multi-threaded bidirectional socket implementation (both examples of things I did recently with DeepSeek flash).

Angular can only be written in one way so the UI was trivial. I know the architecture for the socket implementation and the trade offs for various approaches, so I sketch out the implementation and get DeepSeek to complete it (error handling, keepalive messages, timeouts, etc).

I don't think I would have saved much time if I just asked Fable to "make the socket implementation" but even if it got it right the first time, it would have only saved me a few minutes given that's how long it took to write any way.

Even in more conventional applications, Node.js / React CRUD applications - "write a graphql query for blah" "add an endpoint to run query" "add validation to endpoint" are all trivial for DeepSeek flash. In most of these cases, I have found you're constrained by context window size because these are rarely well architected applications.

[overfeed]

> Experience allows you to design the skeleton where the implementation details are often inconsequential.

Expanding on this thought a little more: it is possible to set up scaffolding that make incorrect implementations inconsequential. If the LLM can detect when the implementation is wrong, it can retry with the errors feeding back into the loop. This is shored up by up-front investments in tests, API definitions, strong types, linting rules, etc. the various cheap, fast Flash models do not need to 1-shot solutions if capable of autonomous reiteration.

[apatheticonion]

This is a great point. In the past this is how I learned to architect software in general as you are never the only contributor and often must accommodate the contributions of contributors who are early in the careers. Good software architecture reduces the damage possible by low quality contributions.

I have had great success with contract driven development workflows in this regard.

LLMs being effectively very knowledgeable jr engineers, slot right into those patterns.

Another interesting thing is that Rust _really_ shines with LLMs. The compiler gives very helpful messages which the LLM bounces off very effectively. If it compiles, it works, you just need to double check the logic is correct.

By contrast, with something like Go, you'd need to be on the lookout for nil pointers, race conditions and so on. Node.js / TypeScript is less dangerous, but you still have memory leaks that aren't possible in Rust and lack pattern matching so must be vigilant about type discohesion.

[hodgehog11]

This is mostly true and a month ago I would have agreed wholeheartedly. However, there are still situations where almost every line of code has to be deliberately chosen in a delicate way. Think research code or settings which demand a high degree of optimization. Opus or GPT-5.5 are absolutely miserable to guide in these settings.

"Well, don't use an AI for those purposes!"

That used to be true (and is now again true I suppose). Fable was a peek at something different, it seemed to be actually able to start to tackle these kinds of problems. That saves a lot of time, since checking these programs is often far, far easier than writing them. Experience is needed to check and scaffold still, but something like Fable becomes a prerequisite for these settings.

[iwontberude]

People are wasting tokens with perfectly fine models by not planning effectively. Anyone who says Opus is not capable of doing what Fable can probably are not planning effectively and just praying to the one shot gods.

[overfeed]

> At least on interesting problems.

What fraction of your work is "interesting problems", and what field do you work in?

> If you’re doing generic web dev work, probably not the case

I have a feeling the bulk of most people's work is "generic $X work." My take is people should figure out their mix of interesting vs boring work, and optimize accordingly. Flash models also tend to much faster,

[zrn900]

Yep. After moving Deepseek for a long time I also moved to Mimo 2.5 pro. It's not only similarly cheap, but it's very good with 1m context window and gigantic tasks. Produces very lean code.

[theanonymousone]

Yes. Whenever I tried, I found no reason to use DeepSeek v4 Pro versus MiMo v2.5 Pro, price-wise or performance-wise. Benchmarks also confirm this.

[trollbridge]

It is since they adopted DeepSeek’s pricing, although it’s not as good about caching as DeepSeek.

UltraSpeed will change how you think about agentic coding.

[alfiedotwtf]

Sadly also one of the slowest of the recently released large models

[NicoJuicy]

I have the feeling MiMo v2.5 Pro uses a lot less tokens and so a lot cheaper ( comparable pricepoint)