HN Simulatornew | past | comments | lists | submit | preommr's commentslogin

> made a HUGE mistake by not publishing a few months ago

Oh no, they'll lose out on some minor press.

Their work will still stand on it's own and for the people that it matters to, they'll know of him and his work.

The broader public probably wouldn't care much anyways, and solving math problems is going to lose a lot of glamour now that llms can do it to large degree. Not to say that it isn't impressive, but there's a large segement that just wants problems solved and doesn't care if it's done by human or otherwise.


I have to agree, those results are surprisingly bad and I am not seeing an improvement.

For sol 6.2, some of them are slightly better at higher cost, some of them are way worse and only slightly cheaper (but comparable to terra in quality while being more expensive).

I guess we'll see as more information comes out, but so far not looking good.


I care less about the models being good and more on if I can fallback to them if Opus 5.5 gets a massive price hike after the IPO

This is what happens when you burn billions for no reason.

Github copilot had INSANE usage. Cursor switched away from per request plans in 2025, Copilot did it this year with all the reasoning and more expensive models. It was insane, you could write out a whole page of features, and have it run on one request for hours depending on the loopholes you exploited (some models were discounted for some time, basically free in some cases, and shenanigans around switching in betwen agents)

We talk about opus et. al being heavily subsidized at like 1-2k for $30-$100. I remember seeing figures in the 10s of thousands in subsidizations for GHC. There's no way they weren't losing just massive piles of money.


> More generally, the problem is almost always motivation and discipline, not the availability of instruction.

This is completely wrong in the context of AI.

Lots of learning can be massively improved if someone can zero in on what specific misunderstanding someone has because of how they interpreted some information. You might have the books, but it doesn't help if it presumes you understand some other fundamental hidden in another book.

You can't constantly ask even direct teachers about these questions, because they're paid by the hour, and just asking the questions can take a quarter of an hour. The best mitigtation was basically having an excellent lifelong education at every step to make sure the gaps were mostly filled in and students ask the occasional question for the limited time teachers can afford them.

AI changes this completely. AI can teach you exactly what you need all the way back to the exact fundamentals, correct any misunderstandings, poor wording, show you 10 different examples until it clicks in the way that's best for your brain, draw diagrams and so much more.

And it does this for pennies, without judgement and lots of patience.


Do you have kids?

They have either never seen a kid who doesn’t want to be there and won’t ask questions, or they imagine that the sheer excitement of AI would overcome that, it sounds like.

It’s certainly not doing it for pennies at Alpha School. And all of that is what a personal tutor can do already. For $75k you could hire a professional teacher to spend two hours/day teaching your child. Now we just have to have someone run a comparison, maybe find a group of parents who wanted to enrol in Alpha School and divert them to this plan?

Well if the current AI leaders are to be believed, it won't matter because money won't be a thing, robots will do all the hard labor, and cancer will be solved.

All in the next few years btw, around the same amount of time to actually build the factories. So strap in I guess.


I just don't see how they don't get that valuation.

Not because it's sensible (it isn't), but because of how much hype is behind it, which is all that matters nowadays.


It's always been based on confidence/emotions

> No one will lend people money to build it

The money is there but not at the risk levels these companies have pushed things to. They're spending billions like it's candy, and treating millions like change they found in their couch cushions.

Financial systems are usually in the hands of people that have some risk management. When these labs say things like this will change everything, there's no way to know how to evaluate any of this.


Feels like ai image generation has plateaud.

Coding related have had massive step-up improvements over time.

The same cannot be said for image generation. If anything, I swear it's gone backwards.

It peaked with very early midjourney (like 2022), maybe follow ups with stable diffusion, and then with some of the nano banana models for actual editing. But overall, very lackluster in comparison to other areas.


There have been a lot of strides in open-weight models, well beyond SDXL. Flux, back in August of 2024 was a game changer, finally allowing us to move past SDXL’s nuclear pile of comma-delimited keywords for prompting and allowing more natural language.

And this summer we got Ideogram V4 and Krea2, both of which are highly capable. You can see how the open/closed models have advanced since SDXL in the Progress tab.

https://genai-showdown.specr.net

MJ is pretty out-of-date by this point though it can still be fun from an image space exploration perspective. It scored a rather pitiful 2 out of 15 on my bench back in the summer of 2025.


Progress peaked in 2022?

Otherwise you’re implying the quality plateaued 4 years ago which is ludicrous.


> Feels like ai image generation has plateaud. > Coding related have had massive step-up improvements over time. > The same cannot be said for image generation. If anything, I swear it's gone backwards.

I would argue both are exactly the same. The images are getting nicer, the AI can alter pictures better. It's the same as coding.

The difference is our perception. Looking at image takes 2s, and we can spot what we don't like. Looking at a code change takes hours to understand. It's very difficult to tell how much code generation is improving because almost nobody is reading it.

Image and code generation don't correlate exactly, but I would be surprised to see a big jump of capabilities in either without the other quickly following.


Code's easy to evaluate because there's finite number of things it needs to do and its end goal isn't to be perceived by a person. An image can always be wrong.


Your feeling is generally right. There have been no seminal moments in image generation anymore in the last couple of months. Most attention from the big labs has gone to video since that‘s where the money is. We have a page to track the history of it all here https://lumenfall.ai/arena/history

(Right now missing some models which I‘ll add today.)


You could argue that progress has been relatively flat since Nano Banana Pro, but that's still less than a full year old (just barely), and that was the first model that could really fully maintain visual consistency across complex edits. The GPT ones are a little stylistically disappointing in ways, but they've still been making steady progress too. I don't think we can claim a peak just yet, even if it's been slower.


Try running Krea 2 locally (I'm using ComfyUI). If you think stable diffusion peaked in 2022, I think you may be a bit surprised at what it's capable of. It leaves modern competitors like Flux, Moss, etc in the dust.


Well Moss just came out last month so it's a bit early to dismiss it. While Krea2 is excellent, I think most people who use this stuff on a daily basis wouldn't necessarily agree that Krea2 is leagues better than Moss.

In my testing, Moss has slightly better prompt comprehension and adherence. Krea2 seems more suitable for photorealism, while Moss tends to be a bit better in other areas like illustration.

Another thing to consider in Moss’s favor is its more permissive Apache 2.0 license.

https://genai-showdown.specr.net/?models=k2,moss


I disagree entirely. Imagine generation in the last 6-months has been an entirely different animal. Generating large artistic concepts has plateaued for sure, but nitty gritty image work is massively better in 2026.


I think with software engineering the biggest jumps have been agentic platforms or work flows. You can create a ticket and this triggers na agent to read the relevant docs and submit an MR, another to validate multiple concerns (correctness, security, performance, costs), another to uodate docs, etc. That's far more impressive than "dev sits down and asks LLM to write code".

With images, I don't see such potential. A builder and a reviewer iterating on improvements? Maybe the reviewer noticing it just needs programmable improvements, eg contrast, color balance etc and triggering a script on top of the LLM generation?


There’s less money to be made with images, than code gen.

Investment in new models follows the money.


Coding flattened out like 9 months ago. In reality, people are getting dumber and lazier from over using them, so they feel like they're improving when they aren't really, outside some areas like RE.

In reality RE has only gotten better because they RLd it into the models because they NEED data, so now every kid is decomping proprietary softwate and sending them back great training data.


>Coding flattened out like 9 months ago.

It absolutely has not. You do have a nice point with data though, somehow haven't thought about that


> Coding flattened out like 9 months ago

Opus 5.5 begs to differ


Even on benchmarks it hasnt gone up that much. Ya'll are extermely one shot by marketing and twitter posts of one page demos.

Not everything runs in a browser, it still creates fires in codebases and generally produces spaghetti code. It gets the job done for 50th percentile problems for sure.

Turns out not everything is a 50th percentile problem.


You know most of us are using these models day-to-day, right? Opus 5.5 is leagues better than Opus 5. Hell, it's better than Fable 5.1.

I have to imagine you don't if you can say this with a straight face...


> Feels like ai image generation has plateaud.

I'd bet the future is in combined AI-driven 3D modelling + rendering + AI post-processing. This pipeline solves quite a bit of typical image generation problems.


I have had some success recently, generating concepts in MidJourney, and then getting Opus recreate in 3D. Definitely a workflow I will be experimenting more with.


Krea has kept improving editing


Seedream 4 continues to be the best model by far at character consistency. I havnt done my tests on this model yet, however.


"It takes 90% of the time to do 90% of the work. The other 10% takes the other 90% of the time"


Turns out you can't brute force art, to the surprise of no one except tech bros.


Visual creativity requires experimentation - trying things that shouldn't work, and teasing out some measure of emotional or intellectual interest. Systems that are increasingly only capable of outputting marketing visuals are caught in a blandness spiral. They're just going to get worse, so long as they filter out all the weird things artists might actually want to generate (six-fingered hands, off-center subjects, nude bodies). Imagine training an artist solely on magazine ads, or only allowing them to produce that kind of work. No figure drawing, no sketching from life where a pen slip produces a unique texture that becomes a signature element.

I say this as someone who's generating smut-adjacent, art-adjacent images almost every day. It's amazing what image models can produce, but (as with my traditional skills) they'll always be held back by a refusal to embrace holistic study of the world. Sometimes you gotta draw a lamp or weird sculptures or a vacuum cleaner or a dick. Sometimes you have to paint an empty sky or a nipple and do your damnedest to find an epic novel in them.


Bitter lesson learned in 3... 2..

I know Dylan Field recently posted they made the Llama's 70B model do amazing things but the AI training field is an absolute snake pit.

But then again they have money to burn ig.


You've misunderstood the bitter lesson if you think it applies here. It's not "frontier LLMs are always better", it's "algorithms that scale with compute/data will usually end up surpassing algorithms that are based on human understanding".

Fine-tuning an open-source LLM with reinforcement learning in order for it to be better at a specific task does not go against the bitter lesson at all.


> You've misunderstood the bitter lesson if you think it applies here.

That's fair because I am stretching it.

I think there's AI companies that understand and scale AI. Then there's the rest that are doing a modern version of writing domain specific knowledge using skills, and maybe a large collection of their own data.

Maybe these focused models outperform in their own niche similar to how hand-written rules outperformed in their own niches. But long-term, they get outpaced quickly by generalized approaches.


Guidelines | FAQ | Lists | API | Security | DMCA | Apply to YC | Contact

Search: