Comment by cultofmetatron

The streetlight effect:

> A policeman sees a drunk man searching for something under a streetlight and asks what the drunk has lost. He says he lost his keys and they both look under the streetlight together. After a few minutes the policeman asks if he is sure he lost them here, and the drunk replies, no, and that he lost them in the park. The policeman asks why he is searching here, and the drunk replies, "this is where the light is"

All of your suggestions are better but they're hard, so someone casually evaluating an AI isn't going to do them.

One-shot performance often translates to the most difficult problems a model will be able to understand. We run an evaluation that tests both agentic and one-shot performance, and we find that Chinese models are almost universally very good at using tools and a harness to iterate towards a better solution, whereas their initial response ranks relatively low.

Compare that to Gemini models, which have impressive fluid intelligence on the first response, but fail to call tools or explore correctly which limits their usefulness for agentic coding.

Neither will be great for coding in a computational chemistry repo for different reasons, but the model with strong one-shot performance will be less likely to make subtle errors indicative of poor understanding, so we weight both capabilities into their final score.

The latest Anthropic and OpenAI models excel in both domains.

Data at https://gertlabs.com/rankings

IMHO, It's not the oneshotting.

It's the "starting from empty slate" greenfield that's the real problem.

We used to make fun of Engineers who follow a README on a framework, test it on an empty project, and say "this framework is the best for our 10 year running production app". Greenfield mentality is always the solution to all problems and problem to all solutions.

One should still measure oneshotting, it's an important self-measurement metric - but against an established, large codebase.

It's a proxy for what you actually want to measure.

Note that after the model generated a bunch of (intermediary) code, they still have to have it tested and get bugs fixed (via the agent/harness). In this "one shot" you still have agent loops against human defined objectives.

And these toy examples give some insight as to how the model performs. If the test were "here's some code written by $corp, please take these tickets and work on them" it may be a "real" example but nobody would be able to make sense of actually how "hard" it is, or how "well" the model did the job, besides the workers already familiar with the context.

At least everyone knows what a 3D game is.

What are you yapping about? This was not one shot prompting, but a long run horizon task. But GLM and Opus invoked at least 120+ tools across the runs.

I guess the experiment is interesting to determine if a model can produce something subjectively valued as "good" based on fairly vague and open-ended specifications. The benchmark is not to determine if the output fits the input, but whether the output is internally consistent: it's a game, but does it behave as one would expect that any game behaves? Does it end when you each the goal, do you die when hitting the spikes, are there weird edge cases in behavior when you move around?

I think however that they should have used the same harness and also repeated the experiment a few times to judge the variance in results.

> I'd rather see a coding agent that can follow steps in a plan file to a T while following guardrails and adhering to the proper coding conventions in the human reviewed spec.

In fact, I'd rather see Anthropic publish a convincing project that does this using Claude. The project should be complex enough and novel enough to show the world how reliable and powerful Claude is. That is, Anthropic does not need Amodei or its employees to tell us that whatever percent of engineers will lose their jobs. They can just show us. Easily.

It's true that no one is trying to one shot anything serious right now, but it's still an important metric. Claude Code and Opus really took off when they improved the harnessing enough that it would self-correct many of its mistakes without needing user input. In fact I think long-term autonomy (in the range of several hours) and self-correcting is going to be where we see most improvements in coming years.

Unless I'm missing something, the prompt he gave must have been fairly detailed because both games are basically identical.

But for a more practical issue, the ultimate goal of LLMs is to replace software engineers, or at least enable everybody to become a software engineer, to use a more up-beat phrasing that's no less accurate. And so an LLM's ability to reliably construct something from a poorly defined, contradictory, or otherwise flawed prompt, while accurately inferring intent is probably the first finish line.

I feel like on HN there is an endless cycle:

- Vibes are too subjective, I want an actual A/B test!

- An A/B test is too limited, I want a benchmark! (You are here.)

- Those benchmarks never seem to be reliable, I just go on vibes.

Exactly this. I recently tried Claude code again to get the subsidy on fable rather than paying api prices and was so frustrated by how much it pushed autonomous behavior. It would start ignoring my planning documents, ignoring my coding conventions, reimplementing features and code already in the project (not sure it ever makes sense to have two auth systems in parallel or two websocket implementations for the same ui) and then in the most shocking interaction just refused to stop working and listen to my instructions. I think maybe it was because there was a subagent doing the work but it was a complete waste of time and effort.

I was using cursor, in large part because I could at least stop it when I need to.

I ended up building my own IDE from scratch so I can be more in the loop while also having the full agent experience.

Isn't a plan file just a single long prompt?

> I'd rather see a coding agent that can follow steps in a plan file to a T while following guardrails and adhering to the proper coding conventions in the human reviewed spec.

Guardrails/conventions should be enforced in linters, formatters, static analysis tooling; not specs/prompts.

> I seriously dont' know all this big hullabaloo about one shot prompting.

It's a relatively objective way of testing LLMs, and I think it's pretty representative of how strong models are overall.

The outcome of this test mirrors how GLM 5.2 and Opus 4.8 work for me: they're both similarly capable of fully executing a given task, but Opus tends to have a bit more "taste" in how it handles unstated details or implicit requirements.

> what you'll get is a series of assumptions made by the model

Yes, but that's why we use these models in the first place. We don't want to explicitly write down all the details because that would mean writing code. So we write a higher-level, human-language spec, and let the LLM fill in the blanks. The question is how good they are at doing that.

One shotting is useful to test but only with a huge prompt (eg, build something according to this spec).

I agree generating millions of tokens from a handful of input tokens doesn't convey anything meaningful to me.

If a model can take a series of increasingly complex instructions and satisfy the requirements without human intervention, we can pretty easily decide how well overall the model does. And, judging better models just means adding more requirements to a task. So, I think it's a useful method (Even if it's not a realistic use case).

Of course, with a software engineer at the helm - the models are going to be able to be guided to produce much better output. (Or worse, depending on the engineer!)

I think you're underestimating the elegance of "hey build X". It already captures a lot of what you're interested in.

Additionally, with "Hey build X" nobody is happy with the methodology and people rightfully complain about the set up.

Using your suggestion the methodology would require a lot of presumptions & arguments regarding why you choose it and think it relevant to people.

Either people would not "get" it quickly enough or would disagree/not be interested on the setup because its not how they use LLMs.

On one hand, that's sort of true for practical uses - and benchmarks notoriously undercount multi-turn settings.

On another, being able to reliably tackle minor tasks with no handholding is very valuable in itself. Sometimes implementation details are important, but often, the most important thing is to Get It Done.

When the model produces reasonable results from one prompt, you could assume that it will also return reasonable results through the follow up prompts.

The argument is flawed, there is no logical reason to assume a single prompt won’t be sufficient to constitute the complexity of a software project. It may not be practical in many cases but there is too much variability in what is considered a complex software project and in the sufficiency of instruction in a single prompt to make that claim and say it’s “by definition.”

> a single prompt wont' constitute the complexity of a software project.

The top agent is for steering, but all subagents are mostly oneshot prompts

I also love the term zero-shot in the AI benchmark world. So logical. So intuitive.........

"We did multi-shot prompting to try and get these two games into comparable states using these two different models."

"Well obviously you provided better follow-up prompts to the one that came out better."

Also nothing about human-provided plan files and guardrails preclude the one-shot benchmark test. Heavens, I almost said "real coding," but in "real agentic program creation" you'd obviously be doing multi-turn interaction with the agent, but how can you provide a fair test when the model's output n determines your n+1 response?

Blame anthropic, they decided to make these type of one-shot examples the primary focus of the Fable 5 release, and relegating benchmark scores to the pdf.

PREACH. I have no idea why THIS has become the standard for illustrating model capabilities. It's endlessly frustrating when that was the initial objective for all these models, but, became increasingly clear over time that none of these models were ever capable of getting the desired output for complex software on the initial prompt.

The reality is: - business rules change - ideas for improvement may arise from the initial prompt - updates to submodules/functions/configs/secrets are BLOCKERS ... etc.

One shot prompting for the expecations of complete software is seemingly more and more a show of incompetence of the use of this technology. It's like trying to make my toddler eat a ham sandwich from the peanut butter & jelly I put in front of him.

That's precisely the difference between an engineer and a business guy.

The business guy would say "hey build me this and that" and would get _something_ to show of.

An engineer will have a long conversation with a llm about the exact requirements, tech stack, tradeoffs. He would understand what is built, how is it built, and refine on the fly until he gets something sensible.

It won't be as fast as "build this", but the result will be much better and more maintainable.

For the enginering workflow, you don't need Fable. Any model better or equivqlent to Sonnet 4.6 would do. Yes, sometimes it will hallucinate, sometimes it'll be wrong, but it's our job as engineers to correct it and have full ownership of the result.

Single prompt performance is interesting because best agentic results of yesterday turned out to be best single prompt results of today.

If we stopped developing LLMs the the only reasonable way to benchmark them would be to compare yheir performance with all the tricks we can build on top of them. Sine the are still developing rapidly any apples to apples comparison is worthwhile.

Of course this particular benchmark is not really single prompt but rather "agentic without steering".

I think that’s the point of the Superpowers SKILL

The thing with one-shot prompting is that it tests the ability for the model to make good choices on its own, rather than only instruction following.

Instruction following has been down for years, and while there are of course metrics that continue to improve as the frontier advances (for example, the ability to continue following the original instructions even as context grows), you can't really get that much better at performing a list of instructions as-written if the instructions are sufficiently precise enough that there's no wiggle room for interpretation (which seems to be what you are describing).

For example, one of the things that got me the most excited for Fable 5 was its ability to work for over eight hours straight on a single instruction and seemingly faithfully the entire time. That was something I observed personally after trying out the same workflow that runs for maybe two or three hours with Opus and then still needs followups. Fable needed no followups. That's a game changer for me compared to the prior state of the art.

That kind of stuff is going to end up being the most beneficial to people who are touching the edges of their knowledge or even exploring completely new areas. And that type of work is exactly the kind of work that makes agentic coding so powerful, even as much as it gets harder to judge the quality of the work when you lack the skills yourself. It's a good thing that the quality increases across the board, even for skilled practitioners.

For example, even people who know how to write inference engines or how matmul kernels work or how to optimize model architecture can't always predict just the sheer breadth of things agents can try to improve performance, and sometimes you get over some wall and reach a completely different optimum that you just wouldn't have reached in any reasonable amount of time by applying traditional knowledge even if you're an expert in the field.

That kind of stuff is amazing. And that's exactly the kind of stuff that one-shot prompting is testing for. It's kind of like testing for the model's "innovation", as much of an oxymoron that is.

Yet this is how virtually everybody is benchmarking and fine tuning.

Since Opus 4.6 I've seen later Anthropic models being more and more capable on one hand, but also less useful on multi turn open tasks.

It feels like with each model they are more and more prone to go "their own way" and jump into the implementation as soon as they can.

I can't but blame it on benchmarks and fine tuning around prompt-to-solution work.