Hacker News Reader

Comment by dalemhurley | Hacker News Reader

OpenAI is killing it now that they are more focused. Killing projects like Sora et al have seen it go from irrelevant to level footing with Anthropic.

Sol is so much better than Fable 5. Then we get Astra (yet to use it) few days after Fable 5.1 (which is very impressive).

Codex is slightly better than Claude Code.

Good on Sam Altman getting back to basics and turning OpenAI around.

View on HN

I think it mostly shows that there is no moat and the only advantage the U.S companies have over the Chinese is more compute. Qwen Max, Kimi K3, GLM 5.3 are really close to Opus/Sol/Fable/Astra and they are open weights.

they don't have moat in hardware either

Chinese counterpart like CXMT and Huawei is begin producing their own chip

You cant block an entire nation level effort with tariff

I think the moat that China has is energy costs. It's taking learnings from the Bitter Lesson. If you role up scale and compute to the next level, it's energy resources. China has it and sharing open weight models is an effective means of removing the tech moat. This idea has been floating around for a bit now (I'm not taking credit for it).

They also benefit from the commodification of software/knowledge work since they own manufacturing

It's not energy costs. The US produces about 70% more electricity per capita. Chinese households do pay less than half what US households pay for electricity, but that's because the NDRC sets prices below costs for households. They make it up by charging industry more, and the industrial electricity prices in China are roughly 34% higher than in the US.

> The US produces about 70% more electricity per capita.

And consumers use 4x as much per capita. Industrial generation per capita China comes out ~2x

> industrial electricity prices in China are roughly 34% higher than in the US

For which industrial customer and where? Chinese compute hubs are on par to slightly cheaper on pure electricity costs.

Conversely the US makes it more expensive with interconnect and upgrade fees as well as hefty take or pay contracts.

A 1GW datacenter in VA for example would add 5-10c kWh and a 12 year take or pay deal

per capita seems the wrong metric given the difference in population sizes and America's wealth. They have roughly 4x the people and have added 10x new power capacity in the last 10 years, not to mention lapping us in renewable and long distance transmission lines added.

If there was no moat, nvidia and meta would have SoTA models too.

Meta is awfully close.

lol! Good one...

Went from years behind to months pretty quick.

It is not in nvidia’s interest to be too good at model creation

Why not? Commoditize your complement, and all that.

And if they get too good, they risk harming or otherwise killing their golden geese (their customers), who they are heavily invested in.

How? Imagine an open-weight model comes out that is somehow better than proprietary solutions. Now the marginal cost for the consumer is just the cost of renting the inference hardware, without having to pay the overhead of the owner of a proprietary model. And because it is cheaper, more customers want to use it, and Nvidia will sell the providers the inference hardware that they need.

1. No open ai and anthropic means no buying gpus to train. Now nvidia spends money on hardware training their own models. Opportunity cost plus expense.

2. Any open models created from this will not necessarily need their silicon, see apple mlx.

1. I don’t think that’s a very strong argument. OpenAI and Anthropic don’t buy the vast majority of GPUs they use they rent capacity.

Nvidia could just the same rent those GPUs out for inference and actually have way better margins than they do right now. Antitrust and putting all your eggs in one basket are why they don’t, similar to TSMC.

2. Neither do AI labs. See Anthropic buying TPUs, deploying with AMD. OpenAI on Maia, Cerebras, their own wafers.

It’s not about whether or not Nvidia will be able to sell hardware to these providers - it’s about literally killing companies they are financially invested in.

Why would you invest money in a company, and then enter the market to compete with them?

But it is in their interest that their customers can use their models as a base for post-training and LoRAs.

They don’t necessarily need their own models for that

They have models for that. That's what the Nemotron series is. Not just open weights but open training data too and full tutorials on how to use them to fine tune or train your own models.

They exist to keep people using and advancing the tools on their hardware.

Nvidia does have one of the best completely open models. Open weights are nice but Nemotron is open training data too.

bringing the price down b.c. competition != no moat.

There's not 100 frontier labs, it's not like airline companies

About the same, 5-10, when you consider major (aka frontier) airlines.

Actually not a bad comparison. Both burn massive amounts of up front capital to protect an oligopoly in the hopes their commodity product eventually pays off.

The "moat" is the "harness", the app.

For most people, the app IS the AI.

And even for its wonkiness, ChatGPT has had the best UX/UI of them all.

The way to win the AI wars in the eyes of the common folk is through the frontend, to be the Apple of AI, as it were.

this basically says you don't believe there is real AI.

Read the second line guy

There are people all over the world who have no computer skills but they use ChatGPT on their phones daily

They don't know/care shit about models and all that

For them, if the app sucks, the AI sucks.

They have a lot of moat, i'm not sure what youa re talking about. Only amatures are using Qwen, open source stuff that is 3-8 weeks behind. Plus OpenAI has some verticals that keep people in there.

In what way do they have a moat? A cursory look at https://artificialanalysis.ai/models/gpt-6-astra#intelligenc... it lands at 61, only a single point above glm 5.3 while costing significantly more.

The only moat they appear to have is by hoarding compute, and the current trajectory of hardware shows that isn't permanent either for very long

I wish people could see how some of this reads. You are an “amateur” using a model 6-8 weeks behind? Really? Sigh.

  I think it mostly shows that there is no moat

You can argue that TSMC has no moat since Intel and Samsung are also able to eventually make a node as good as TSMC - just a few years later and at smaller scale.

And no one would say that about TSMC.

So there is clearly a moat there somewhere.

Yeah, I'm not sure if "no moat" analogy stands for chip manufacturing. Even if foundries acquire lithographic nodes, the procedures (temperature, duration, etc) are for them to figure out and are usually kept secret. This secret could be the "moat" that differentiates each foundry's operational capabilities.

No. In the semiconductor industry, the "catch-up" player isn't normally spending less in absolute R&D terms.

Comparing the R&D costs of creating GPT-4o vs. DeepSeek V3 (the latest gen for which we already have good accurate numbers) it looks like the latter cost 1/20th as much to create.

If Samsung could catch up with TSMC for 1/20th of the cost, people definitely would say that TSMC has no moat.

Why do you think Chinese models cost 1/20th to train?

That's the ratio the widely published numbers give [1]. One does not have to believe the numbers [2], but those who do believe them are then justified to conclude that there's no moat.

Which numbers you believe is of course going to affect whether you think there's a moat or not. That's largely orthogonal to your TSMC/Samsung analogy I responded to. If you think the "moatists" are wrong because they believe the wrong numbers, that's fine, but then there's no need for the analogy.

[1] https://galileo.ai/blog/llm-model-training-cost

[2] https://medium.com/@theiand/how-can-deepseek-a-5-6-million-l...

But fundamentally, why is their cost 1/20 and is it sustainable in the next 10 years of competition?

Now that is a good and interesting question! Hopefully a "no-moatist" will share their reasoning.

Ultimately, that's what I need to be convinced. No one has put forth a good argument yet.

Clever architecture --> Ok but OpenAI/Anthropic can use these as well and they also have very smart people with their secret clever architectures

Distilling --> Ok but distilling means you will never be smarter than the original. Furthermore, reasoning is now hidden by private labs and they have poison pill answers for distilling if they can detect it. They will be able to detect distilling better and better.

Cheaper electricity --> Ok this is cancelled out by their chips being much less efficient due to not having ASML EUV machine access.

So I don't see why fundamentally their training costs are cheaper over the long term.

I'm looking for a no-moatist to convince me.

Labor. Smart labor would be much cheaper I'd reckon in China than in the US.

How much advantage in costs? What % of labor is training cost?

Mercor, Tacit Labs, Handshake AI... I suspect companies like these play a big part in model improvements, generating high quality benchmark/task-focused data for training.

However, these do require educated, white collar, workers.

Considering that frontier scientists and engineers in the US are currently taking home seven (or even eight, in some cases) figure salaries - pretty high, I'd reckon.

Would like to see the math since the claim is made.

I am not a "no-moatist" per se but one can argue their might be a plateau to how good a inference llm can become. If this is the case the playing field shifts to context, tools and harness, which are much cheaper to build an compete on.

Because they're distilling frontier models and that's a lot faster and cheaper than training a frontier model from scratch?

So why can't OpenAI/Anthropic also distill the good parts of free Chinese models? It's even better and easier for OpenAI and Anthropic. No poison pills as well.

Please just say what you want to say.

From my experience with complex coding tasks (AI infra), I don't think these open weight models are close.

Not sure if I agree, I tried GLM5.3 and it was pretty decent. Ok, it's not Opus, but maybe it's Sonnet?

> Sol is so much better than Fable 5

I'm genuinely so confused when people say this with a straight face. Are you talking about coding? Desktop use? Prose? Or something else?

Sol is a much smaller models and it shows. It often misses the forest for the trees.

>> I'm genuinely so confused when people say this with a straight face. Are you talking about coding? Desktop use? Prose? Or something else?

Same. It makes me wonder what types of things the person must be working on.

This is perpetually an issue with the whole field of AI/LLMs. The experience is so personal. Every time I talk to someone about their use of LLMs for software engineering, I'm shocked by their approaches and experiences. They say "X model keeps missing things" when I rely on it heavily for being thorough. They say "Y always gives me the best results" when I can't stand it.

People will see/think that I'm doing very well with my LLM use, and ask me what I'm doing. I tell them, they try it, then later they come back to me saying they just couldn't get it to work.

It’s really inconsistent. There are sessions where it nails everything perfectly and I leave happy. Then there are sessions where every turn it corrects itself and changes it mind. One session recently I found it funny how every single time it did this one task it tripped over itself and killed its own connection. Like 20 times. It didn’t bother me I just found it odd how despite it being noted down in its state file it kept doing it over and over like some idiot. Literally they can’t learn from their mistakes yet.

This is the job now, we are shepherds.

This is why I generally don't trust benchmarks, or anything other than my own experience tbh. It always seems like everyone has a different answer.

If we truly had some AGI model, it would probably be fairly obvious to us all no?

I feel like a lot happened this week and people are glazing how ridiculously strong Flash 3.8 is right now compared to Fable/Opus/Sol/Astra.

Flash 3.8 is rad. Easily my daily driver now. Only downside is it's Gemini so sometimes it just keeps going until it wants to be done.

I have a few attention and finish mechanisms in my prompts. I have been using it for a week and a half and with some prompt taming it is great. (I have early access to the models cause I work at the place that makes the model). None of my attempts to ever tame Opus 5 have worked.

I’ve found Sol performance to be incredibly spiky. It has tremendous IQ and can fix very difficult bugs. But it is horrible at design (both visual and system design), anything that involves thinking about users or UX, and massively overcomplicates almost all work.

I noticed the same. I wanted a simple crud webapp and suggested an insane techstack involving C#, Razor Pages, MSSQL and more. I went with my planned setup of python flask with an sqlite db which served me well for years.

It's still incredibly important to have a human in the loop correcting design decisions and having good taste.

> insane techstack involving C#, Razor Pages, MSSQL

Is a very sane tech stack, you're just biased against Microsoft.

Half the world's enterprise apps run on that combination, or a minor variation of it.

Like Java it is full featured ("batteries included") but unlike Java it is relatively terse and actually pleasant to work with.

Oh, and unlike Python, it is very fast, within spitting distance of compiled Rust and C++ web apps.

There are a million and one reasons to be biased against Microsoft, regardless of the fact that C# tech stack is decent

Correct I'm biased against technologies that only run on a single OS for no benefit to the user.

ASP.NET runs on MacOS, Linux, and Windows.

Microsoft SQL Server now (also) ships as a Docker container that runs on Linux.

> Correct I'm biased against technologies that only run on a single OS for no benefit to the user.

Do you ever use software that only works on Linux?

Do you use an Android phone?

.NET has not been a particularly Windows specific framework for almost a decade now.

You could have just added “flask SQLite stack” to whatever prompt you added. Just those three words, randomly somewhere in your prompt.

Was your prompt just "I want a simple crud webapp" and that's the extent of it? There's absolutely no way you included the words "python", "flask", or "sqlite" and it still went with a Microsoft stack.

Dotnet minimal APIs plus mssql is fine for simple crud apps… I would do Postgres, but that’s me.

Swapping mssql to SQLite would also work perfectly

> massively overcomplicates almost all work

People with high IQ often do this IRL. There's training tension in this area. Intelligence and overcomplication correlate and are hard to extricate.

Intelligence is actually correlated with the ability to simplify complicated things. Occam's razor. Compression as comprehension.

We're not asking the model to simplify something, we're asking it to perform a task. Its subtle preferences show up as an overcomplicated path to the goal.

In some cases, there are also nuances that we don't pick up on. Here it's our preference for simplification that's showing up. We set the lossy compression factor higher than it does.

I vastly prefer Sol. It does what I tell it to almost exactly, pretty much every time.

I work on very low level stuff (think RTL/FPGA, firmware, software where optimising for nanoseconds is just normal).

For me Sol is the only cost effective model available. Fable 5.1 is indeed good and vastly better than original Fable (which refused to work on most of my stuff for 'safety' reasons).

It's very good at this sort of low level stuff to the point that I really can't understand/relate to people having a good time with Opus (which comparatively performs extremely poorly on my particular workload).

I also just don't like how lazy Anthropic models are. They will do 10% of what is asked and then summarily declare victory.

Sol on the other hand is more like "one of us", slight touch of the 'tism, extremely pedantic, will go to the edge of the known universe if that is what it takes to prove/fix/build what you asked for or run out out of credits trying.

It's a personal and workload dependent thing. For me right now Sol for 99% of stuff because Fable 5.1 still burns through $5k in credits a day.

Agree 100%. And I also work a lot on lower level / systems stuff (including RTL here and there, too). Opus is sloppy, and leaves negative cases all over. The GPT models in Codex have a more pedantic and detail oriented "personality." Often to a fault.

Sol will leave a mess of excessive redundant tests and isn't so great at abstraction ; but it produces more reliable working systems.

It's kind of nice to have access to both, but I don't have the $$ for that right now, so I just keep the Codex sub

Can confirm this as well, mostly VHDL and HLS. Sol and Fable can reason about performance and designs consistently. Whereas Opus and others seem to just throw generic optimisation techniques at the wall unprovoked (while hallucinating a justification + expected improvement) until the synth reports improve.

Sol better than Fable? What? I've found it to basically be on part with Opus and I max out 2 accounts on both providers every week.

Its funny, my experience with Sol has been awful. It really overworks problems and tracks into areas it does not need to...

I just dont get how its good for some, and bad for others. It makes me suspect that the models performance is not even against problem sets and it really is just a probabilistic prediction machine. Which then makes me very skeptical of GPT-6 Astra, because if their big claim is Computer Use then it is probably bad in a bunch of other areas.

It is funny indeed, people sometimes with same amount of experience with software development, get vastly different experiences from different models and harnesses.

> I just dont get how its good for some, and bad for others.

If I were to listen to my hunch, it would tell me that it's all up to the prompts that ends up going over the wire (including all the bloat some people have), what workflow/process you use and what the existing state of the project is.

You have to bake the 'lazy dev'/'keep it simple stupid' mentality into your AGENTS.md and / or the skills you're using to design things. It will take things too literally sometimes so you also have to make sure you're being accurate. Best way I've found to use it is make it ask you clarifying questions about what you're trying to build and have it help design the shape of the thing. Then it writes the instructions in a format it understands.

I've had Claude do the same thing where it goes off and spends 100% of my tokens on 3 functions and an ungodly amount of tests / scaffolding that do almost nothing when I gave it an underdeveloped idea.

Their ads business is also doing well. Not "will recover all compute costs" well, but crossed $1b in a few months.

This is what Google needs to do and is probably why Demis has stepped back a bit

Codex is missing a few things that Claude code has had for some time like defined plugin subagents and a few other things. But overall it’s fairly capable. The biggest gripe I have is that codex really restricts context window sizes and compaction leads to a lot of grounding work, and overall codex GPT is too literal in many situations - it’s follows direction slavishly, and when subagent reviewers are used, they tend to find increasingly obscure “flaws” on the instruction following impetus, and the harness agent takes them literally as issues to fix even when it leads to bizarre outcomes. For instance I’ve had several runs where it tries to end up building a hermetic system with sha hashing of everything (including operating system binaries and kernels, tool chains, etc) to certify test results are valid, etc. I have to sort of watch it carefully to be sure it’s not drifting into some insane yak shaving corner, which it will happily do for weeks on end.

Claude has the exact opposite problem, especially opus-5, where I literally can’t trust it to print hello world without taking a shortcut, or just simply lying and saying it printed it when it didn’t, behind a giant wall of inscrutable text. I find it very ironic that Anthropic is the vendor of the lazy lying cheating model that does almost everything you tell it to it do.

I’d really kill for something that balances instruction following and loop escaping behavior better. Fable 5.1 does seem a lot better, feeling more like 4.6 behavior, and honestly Sol has improved as well. I’m pretty psyched for the next generation, as I think the competition has heated up so much that things will improve really fast to the point of marginal utility opportunity being increasingly close to epsilon.

You can enable the 1 million token context window and adjust when it compacts in your config.

> model_context_window = 1000000

> model_auto_compact_token_limit = 900000

I believe it does consume your usage a bit faster though.

Unless things changed this used to work but was blocked. You also used to be able to force in a model catalog to get around the token cap. Each, as of at least April, were closed off and ineffectual. I’ve not tried lately so you may be right.

Opus 5 is a genuinely infuriating model. I hate it’s behavior.

We have a mandate to only use 4-8. Sonnet 5 is pretty good and fable 5-1 has been pretty good so far fwiw.

Fable is okay, just slower, eats tokens and not any better at coding tasks. Maybe a little better, but not better enough. It's a lot faster to have a cheap and fast flash agent / sonnet do the implementation work with Fable tagging cleanup and divergence from spec and goals.

Flash 3.8 is genuinely my favorite all around model right now. And yeah Opus 4.6 was the last Opus model I liked. 4.8 is tolerable. Opus 5 is a terrorist. It just can't follow an instruction to save its life and regresses rapidly. Sol at least stays on track so I have to smack it's hand way less often. I am biased, but Flash 3.8 and 3.7 are the first Gemini models I just recommend to others.

5-1 is considerable cheaper I’ve found and feels a lot more like 4-6.

But opus 5 described as a terrorist is being generous.

Amusingly, as an autonomous coding agent, I kind of like Opus 5. But I have to bound it on tasks or it just goes off the rails. But I'm bounded tasks, it is genuinely solid. It's kind of like the new Sonnet 5. Right now my favorite model to interact with on the frontier side is Sol 5.6 so I have been using that as my coordinator. Flash 3.8 is my other favorite just because it is so fast and I use it a lot at work and know its quirks.

Codex's lack of auto-mode is what prevents me from using it for serious work compared to Claude Code.

It has had automode for a bit now. I use it every day at work.

"Approve for me" used to work without issue for me when GPT 5.5 was the latest model.

Nowadays it routinely rejects "git push" to the project's repository.

AFAIK it's just yolo mode, which doesn't actually do any checks like Claude Code's auto-mode (which has a model checking all commands for safety).

No, it says “auto approved” after every evaluated command. Does it work? No idea, I can’t think of a time it ever stopped, but I’m also not ever doing anything dangerous with it. It does use another model to evaluate the commands though.

Put it in an isolated container and set it to YOLO

I don't even put it in a container. I just yolo it. It hasn't done anything bad so far. My logic is that if it nukes my filesystem then I deserve it.

> Sol is so much better than Fable 5.

... looks around ...

Killing Sora was one of the worst mistakes they ever made

please tell us why

Rich media is where all the innovation is happening now and in the future.

Text-to-text is dead, has been since Mistral 7b.

Solved problem (you guys like that one don’t you)

They also demoted themselves from “authority on AI” to “in over our heads” by bowing out in the pathetically defeatist way they did at the worst time possible (Hailuo/MiniMax/Vidu coming up) - they naturally completely missed the wave on audio with random companies like Singify taking that market for free.

They just bowed out. They didn’t try. They didn’t try anything more than baseline text-to-text and they aren’t good at that (or code) either, compared to what others are doing.

It’s a really bad position to be in if you’re trying to be an Apple or Microsoft.

To have a mediocre product and then can’t even serve 75% of the mainstream use case.

Your right

I never used Sora but I recently tried Google Flow and the results are quite good. I have the feeling that Nano Banana 2 has been the image generation champion since it's release so maybe OpenAI feels that they cannot outcompete Google on this task?

GPT 2 is almost always better (80+% of the time?) than Nano Banana 2, albeit slower

On my experiments on GenAI Image Showdown, Nano Banana Pro slightly edges out NB2, though I'm not sure the cost difference makes it worth it.

And in pure qualitative terms, I'd agree that NB2/NB Pro are still on par with gpt-image-2, but gpt-image-2 is far easier to steer and is significantly better in terms of prompt adherence.

100% they should have not given up on video.

That announcement is when I stopped paying attention to them.

Reply on news.ycombinator.com