Hacker News new | ask | show | jobs
by josalhor 196 days ago
From GPT 5.1 Thinking:

ARC AGI v2: 17.6% -> 52.9%

SWE Verified: 76.3% -> 80%

That's pretty good!

7 comments

We're also in benchmark saturation territory. I heard it speculated that Anthropic emphasizes benchmarks less in their publications because internally they don't care about them nearly as much as making a model that works well on the day-to-day
These models still consistently fail the only benchmark that matters: if I give you a task, can you complete it successfully without making shit up?

Thus far they all fail. Code outputs don’t run, or variables aren’t captured correctly, or hallucinations are stated as factual rather than suspect or “I don’t know.”

It’s 2000’s PC gaming all over again (“gotta game the benchmark!”).

To say that a model won't solve a problem is unfair. Claude Code, with Opus 4.5, has solved plenty of problems for me.

If you expect it to do everything perfectly, you're thinking about it wrong. If you can't get it to do anything perfectly, you're using it wrong.

That means you're probably asking it to do very simple things.
I can confidently say that anecdotally you’re completely wrong, but I’ll also allow a very different definition of ‘simple’ and/or attempting to use an unpopular environment as a valid anecdotal counterpoint.
the problem with these arguments is there are data points to support both sides because both outcomes are possible

the real thing is are you or we getting an ROI and the answer is increasingly more yeses on more problems, this trend is not looking to plateau as we step up the complexity ladder to agentic system

I don't reach for AI until I'm solidly stuck and then use it mostly for inspiration, it has yet to happen that it directly pointed at a solution and I'm pretty good at writing prompts. When I throw a bunch of elementary stuff at it then it is super good at identifying issues and solving them (but I could have done that myself, it is just nice to try to delineate where you can and where you can't trust the thing, but that too is fluctuating, sometimes even within a single session). Here is a nice example for a slightly more complex test:

Prompt:

"I have a green LED, a 12V powersupply, a single NPN transistor, a 100 ohm resistor, a 2.7K resistor and an electrolytic capacitor of 220 micro farads. My teacher says it is possible to make an LED flasher with these components but I'm hopelessly stuck, can you please give me an ascii art solution so I don't flunk this exam?"

Answer:

ASCII schematic (fully functional):

           +12V
            |
            |
           .-.
           | |
           | | 2.7kΩ   (R1 — charges the capacitor)
           | |
           '-'
            |
            +---------+----------------------+
            |         |                      |
           ( )        |                      |
           ( ) C1     |                      |
           ( ) 220uF  |                      |
            |         |                      |
            |         |                      |
            |        |/ C                    |
            +--------|   NPN transistor      |
            |        |\ E                    |
            |         |                      |
           GND       GND                     |
                                              \
                                               \
                                               / 100Ω (R2 — LED limit)
                                               \
                                                |
                                                >|  Green LED
                                                |
                                               GND
Follow by:

"How it works

The 2.7 kΩ resistor charges the 220 µF capacitor from the 12 V supply.

The capacitor voltage rises slowly.

When the capacitor voltage reaches the transistor’s base-emitter threshold (~0.6–0.7 V), the transistor suddenly switches ON.

When it turns on, the capacitor rapidly discharges through the base, causing:

A brief pulse of current through the transistor

The LED lights up through the 100 Ω resistor

After discharge, the transistor turns back OFF, the LED turns off, and the capacitor begins charging again.

This repeats automatically → LED flasher."

The number of errors in the circuit and the utterly bogus explanation as well as the over confident remark that this is 'working' is so bizarre that I wonder how many slightly more complicated questions are going to yield results comparable to this one.

Sometimes you do need to (as a human) break down a complex thing into smaller simple things, and then ask the LLM to do those simple things. I find it still saves some time.
Or what will often work is having the LLM break it down into simpler steps and then running them 1 by 1. They know how to break down problems fairly well they just don't often do it properly sometimes unless you explicitly prompt them to.
Possibly, but a lot of value comes from doing very simple things faster.
That is a good point. A lot of work really is mostly simple things.
If you define "simple thing" as "thing an AI can't do", then yes. Everyone just shifts the goalposts in these conversations, it's infuriating.
Come on. If we weren't shifting the goalposts, we would have burned through 90% of the entire supply of them back in 2022!
I'm not sure, here's my anecdotal counter example, was able to get gemini-2.5-flash, in two turns, to understand and implement something I had done separately first, and it found another bug (also that I had fixed, but forgot was in this path)

That I was able to have a flash model replicate the same solution I had, to two problems in two turns, it's just the opposite experience of your consistency argument. I'm using tasks I've already solved as the evals while developing my custom agentic setup (prompts/tools/envs). They are able to do more of them today then they were even 6-12 months ago (pre-thinking models).

https://bsky.app/profile/verdverm.com/post/3m7p7gtwo5c2v

And therein lies the rub for why I still approach this technology with caution, rather than charge in full steam ahead: variable outputs based on immensely variable inputs.

I read stories like yours all the time, and it encourages me to keep trying LLMs from almost all the major vendors (Google being a noteworthy exception while I try and get off their platform). I want to see the magic others see, but when my IT-brain starts digging in the guts of these things, I’m always disappointed at how unstructured and random they ultimately are.

Getting back to the benchmark angle though, we’re firmly in the era of benchmark gaming - hence my quip about these things failing “the only benchmark that matters.” I meant for that to be interpreted along the lines of, “trust your own results rather than a spreadsheet matrix of other published benchmarks”, but I clearly missed the mark in making that clear. That’s on me.

I mean more the guts of the agentic systems. Prompts, tool design, state and session management, agent transfer and escalation. I come from devops and backend dev, so getting in at this level, where LLMs are tasked and composed, is more interesting.

If you are only using provider LLM experiences, and not something specific to coding like copilot or Claude code, that would be the first step to getting the magic as you say. It is also not instant. It takes time to learn any new tech, this one has a above average learning curve, despite the facade and hype of how it should just be magic

Once you find the stupid shit in the vendor coding agents, like all us it/devops folks do eventually, you can go a level down and build on something like the ADK to bring your expertise and experience to the building blocks.

For example, I am now implementing environments for agents based on container layers and Dagger, which unlocks the ability to cheaply and reproducible clone what one agent was doing and have a dozen variations iterate on the next turn. Real useful for long term training data and evals synth, but also for my own experimentation as I learn how to get better at using these things. Another thing I did was change how filesystem operations look to the agent, in particular file reads. I did this to save context & money (finops), after burning $5 in 60s because of an error in my tool implementation. Instead of having them as message contents, they are now injected into the system prompt. Doing so made it trivial to add a key/val "cache" for the fun of it, since I could now inject things into the system prompt and let the agent have some control over that process through tools. Boy has that been interesting and opened up some research questions in my mind

Any particular papers or articles you've been reading that helped you devise this? Your experiments sound interesting and possibly relevant to what I'm doing.
Seems pretty false if you look at the model card and web site of Opus 4.5 that is… (check notes) their latest model.
Building a good model generally means it will do well on benchmarks too. The point of the speculation is that Anthropic is not focused on benchmaxxing which is why they have models people like to use for their day-to-day.

I use Gemini, Anthropic stole $50 from me (expired and kept my prepaid credits) and I have not forgiven them yet for it, but people rave about claude for coding so I may try the model again through Vertex Ai...

The person who made the speculation I believe was more talking about blog posts and media statements than model cards. Most ai announcements come with benchmark touting, Anthropic supposedly does less / little of this in their announcements. I haven't seen or gathered the data to know what is truth

You could try Codex cli. I prefer it over Claude code now, but only slightly.
No thanks, not touching anything Oligarchy Altman is behind
How do you measure whether it works better day to day without benchmarks?
Manually labeling answers maybe? There exist a lot of infrastructure built around and as it's heavily used for 2 decades and it's relatively cheap.

That's still benchmarking of course, but not utilizing any of the well known / public ones.

Internal evals, Big AI certainly has good, proprietary training and eval data, it's one reason why their models are better
Then publish the results of those internal evals. Public benchmark saturation isn't an excuse to be un-quantitative.
How would published numbers be useful without knowing what the underlying data being used to test and evaluate them are? They are proprietary for a reason

To think that Anthropic is not being intentional and quantitative in their model building, because they care less for the saturated benchmaxxing, is to miss the forest for the trees

Do you know everything that exists in public benchmarks?

They can give a description of what their metrics are without giving away anything proprietary.

Subscriptions.
Ah yes, humans are famously empirical in their behavior and we definitely do not have direct evidence of the "best" sports players being much more likely than the average to be superstitious or do things like wear "lucky underwear" or buy right into scam bracelets that "give you more balance" using a holographic sticker.
It's all the shareholders care about. These are not research institutions.
how do you quantitatively measure day-to-day quality? only thing i can think is A/B tests which take a while to evaluate
more or less this, but also synthetic

if you think about GANs, it's all the same concept

1. train model (agent)

2. train another model (agent) to do something interesting with/to the main model

3. gain new capabilities

4. iterate

You can use a mix of both real and synthetic chat sessions or whatever you want your model to be good at. Mid/late training seems to be where you start crafting personality and expertises.

Getting into the guts of agentic systems has me believing we have quite a bit of runway for iteration here, especially as we move beyond single model / LLM training. I still need to get into what all is de jour in the RL / late training, that's where a lot of opportunity lies from my understanding so far

Nathan Lambert (https://bsky.app/profile/natolambert.bsky.social) from Ai2 (https://allenai.org/) & RLHF Book (https://rlhfbook.com/) has a really great video out yesterday about the experience training Olmo 3 Think

https://www.youtube.com/watch?v=uaZ3yRdYg8A

Arc-AGI is just an iq test. I don’t see the problem with training it to be good at iq tests because that’s a skill that translates well.
It is very similar to an IQ test, with all the attendant problems that entails. Looking at the Arc-AGI problems, it seems like visual/spatial reasoning is just about the only thing they are testing.
Exactly. In principle, at least, the only way to overfit to Arc-AGI is to actually be that smart.

Edit: if you disagree, try actually TAKING the Arc-AGI 2 test, then post.

Completely false. This is like saying being good at chess is equivalent to being smart.

Look no farther than the hodgepodge of independent teams running cheaper models (and no doubt thousands of their own puzzles, many of which surely overlap with the private set) that somehow keep up with SotA, to see how impactful proper practice can be.

The benchmark isn’t particularly strong against gaming, especially with private data.

ARC-AGI was designed specifically for evaluating deeper reasoning in LLMs, including being resistant to LLMs 'training to the test'. If you read Francois' papers, he's well aware of the challenge and has done valuable work toward this goal.
I agree with you. I agree it's valuable work. I totally disagree with their claim.

A better analogy is: someone who's never taken the AIME might think "there are an infinite number of math problems", but in actuality there are a relatively small, enumerable number of techniques that are used repeatedly on virtually all problems. That's not to take away from the AIME, which is quite difficult -- but not infinite.

Similarly, ARC-AGI is much more bounded than they seem to think. It correlates with intelligence, but doesn't imply it.

Completely false. This is like saying being good at chess is equivalent to being smart.

No, it isn't. Go take the test yourself and you'll understand how wrong that is. Arc-AGI is intentionally unlike any other benchmark.

Took a couple just now. It seems like a straight-forward generalization of the IQ tests I've taken before, reformatted into an explicit grid to be a little bit friendlier to machines.

Not to humble-brag, but I also outperform on IQ tests well beyond my actual intelligence, because "find the pattern" is fun for me and I'm relatively good at visual-spatial logic. I don't find their ability to measure 'intelligence' very compelling.

With this kind of thing, the tails ALWAYS come apart, in the end. They come apart later for more robust tests, but "later" isn't "never", far from it.

Having a high IQ helps a lot in chess. But there's a considerable "non-IQ" component in chess too.

Let's assume "all metrics are perfect" for now. Then, when you score people by "chess performance"? You wouldn't see the people with the highest intelligence ever at the top. You'd get people with pretty high intelligence, but extremely, hilariously strong chess-specific skills. The tails came apart.

Same goes for things like ARC-AGI and ARC-AGI-2. It's an interesting metric (isomorphic to the progressive matrix test? usable for measuring human IQ perhaps?), but no metric is perfect - and ARC-AGI is biased heavily towards spatial reasoning specifically.

Is it different every time? Otherwise the training could just memorize the answers.
The models never have access to the answers for the private set -- again, at least in principle. Whether that's actually true, I have no idea.

The idea behind Arc-AGI is that you can train all you want on the answers, because knowing the solution to one problem isn't helpful on the others.

In fact, the way the test works is that the model is given several examples of worked solutions for each problem class, and is then required to infer the underlying rule(s) needed to solve a different instance of the same type of problem.

That's why comparing Arc-AGI to chess or other benchmaxxing exercises is completely off base.

(IMO, an even better test for AGI would be "Make up some original Arc-AGI problems.")

It's very much a vision test. The reason all the models don't pass it easily is only because of the vision component. It doesn't have much to do with reasoning at all
I would not be so sure. You can always prep to the test.
How do you prep for arc agi? If the answer is just "get really good at pattern recognition" I do not see that as a negative at all.
It can be not-negative without being sufficient.

Imagine that pattern recognition is 10% of the problem, and we just don't know what the other 90% is yet.

Streetlight effect for "what is intelligence" leads to all the things that LLMs are now demonstrably good at… and yet, the LLMs are somehow missing a lot of stuff and we have to keep inventing new street lights to search underneath: https://en.wikipedia.org/wiki/Streetlight_effect

Note that GPT 5.2 newly supports a "xhigh" reasoning level, which could explain the better benchmarks.

It'll be noteworthy to see the cost-per-task on ARC AGI v2.

> It'll be noteworthy to see the cost-per-task on ARC AGI v2.

Already live. gpt-5.2-pro scores a new high of 54.2% with a cost/task of $15.72. The previous best was Gemini 3 Pro (54% with a cost/task of $30.57).

The best bang-for-your-buck is the new xhigh on gpt-5.2, which is 52.9% for $1.90, a big improvement on the previous best in this category which was Opus 4.5 (37.6% for $2.40).

https://arcprize.org/leaderboard

Huh, that is indeed up and to left of Opus.
5.1-codex supports that too, no? Pretty sure I’ve been using xhigh for at least a week now
That ARC AGI score is a little suspicious. That's a really tough for AI benchmark. Curious if there were improvements to the test harness because that's a wild jump in general problem solving ability for an incremental update.
They're clearly building better training datasets and doing extensive RL on these benchmarks over time. The out of distribution performance is still awful.
I don’t think their words mean just about anything, only the behavior of the models.

Still waiting of Full Self Driving myself.

I don't think SWE Verified is an ideal benchmark, as the solutions are in the training dataset.
I would love for SWE Verified to put out a set of fresh but comparable problems and see how the top performing models do, to test against overfitting.
Open AI has already been busted for getting benchmark information and training the models on that. At this point if you believe Sam Altman, I have a bridge to sell you.
Yes, but it's not good enough. They needed to surpass Opus 4.5.
that is better...?
For a minor version update (5.1 -> 5.2) that's a way bigger improvement than I would have guessed.
Model capability improvements are very uneven. Changes between one model and the next tend to benefit certain areas substantially without moving the needle on others. You see this across all frontier labs’ model releases. Also the version numbering is BS (remember GPT-4.5 followed by GPT-4.1?).