Deepseek 4.1 flash is so fast and cheap. For coding tasks, it only seems to lack the robustness of Opus 5.5 rather than 'intelligence' - it fails to double check critical assumptions, stops before the task is over, and doesn't do enough speculation or exploration on its own initiative. Yet it can do all these things when prompted.

There seems to be another magic ingredient for improving agents beyond model size & efficiency, and so far only OpenAI, Anthropic, and maybe Google know what it is.

Ultimately the question is whether there's a pareto optimal point for intelligence models. There's 2 potential worlds:

1. Frontier models are so good that the token cost of an average task becomes miniscule, so users are incentivized to use frontier models because their sheer intelligence makes them cheaper to use.

2. There is a pareto optimal point where even if you spend more on tokens overall profits decrease because token costs rise faster than revenue does.

The "no lab has a clear lead" world leads to 2.

You bring up a very good point. Not only do lots of people buy inferior products, lots of people employ inferior intelligences

> It's not absurd to think that a company would use its own harness or an open source harness and/or put a proxy in to route requests variably. For two months, they have engineers using different models unbeknownst to them and they measure things like how often the engineers have to correct or redirect the model. At the end of two months, they might find little difference in performance and switch to a much lower cost model.

That's a great product idea!

> But I want them to be valued far more realistically.

They are. That's the problem. The obligations that these companies are, on paper, supposed to meet require these valuations.

If you value Anthropic at $500b that's still a fuckload and vastly less than their obligations and a lot of people are going to say "well wait a minute how am I going to get paid".

The amount of spend going on, and planned, is insane and underpins a lot of this, and is why people are skeptical of the valuations. It has nothing to do with "Wanting" AI to fail and just a reasonable bet based on the math. For them to pay back what they have on their balance sheet entire continents will functionally be a spec in comparison, and that's really really fucking unlikely to happen.

Doubly so when you consider that there's a question of if they can be profitable to projected needed levels without fundamentally breaking some economics parts, and that this looks a lot like the overbuild of the internet during the dotcom bubble, which was also hideously overvalued at the time, for the time frame they were looking at, despite being THE literal world changing tech.

IMO Opus 5.5 has quietly been a revelation. It basically requires no hand holding from me given a sufficiently articulated prompt. At this point I'm limited by reviews and end-to-end validation of changes.

I've found Sol 6.1 a regression over 5.6.

Not due to capability per se, but writing style and polish is much worse.

It will leave really bad comments in the code, commit messages etc without strong explicit steering. Though I've basically moved away from having the model write any documentation at all.

It also tends to talk very circuitously and referencing things implicitly instead of just being direct and clear.

Certainly it's more capable in debugging, but I much prefer Astra over Sol now.

Of course, Astra has its own problems of code golfing everything. But at least it's very concise and direct

Sol 6.1 is definitely not comparable to Opus 5.5 for me.

if you go on short form social media, Opus's ability to do agentic work in things like blender, css animation, etc. have spawned hundreds if not thousands of channels of people pumping out content with it. That and all the video game shenanigans (and now Anthropic is even modifying its system prompt to limit some of these things!)

There's definitely unique things that so far only Opus is capable of doing.

Every time I try and do something with codex/astra/sol 6.1 coming from Claude Code I want to scream. I try it regularly to make sure I'm not getting to tunnel visioned on Claude but at the moment it's night and day.

There's a stock market truism that anyone that has real alpha doesn't sell it or give it away. I feel like there's an analogue here, that if it was just that easy for them to pack up their toys and go conquer the world, I can't imagine they would have been in the business of selling tokens at this point

> Anthropic can remove all access to their LLMs and start replicating entire software companies.

"I have a girlfriend for real she just goes to a different school"

Well what I do know as someone who has a bird’s eye view into many enterprises is that many of them are starting to realize they’re bleeding money to Anthropic and are telling their teams to be more efficient with tokens.

Astra is not the OpenAI analogue to Opus 5.5. Sol 6.1 is.

Wouldn’t previous Opus models be able to solve them too? In my experience Sonnet is often lackluster.

For capital to win they need to get their money out of Anthropic before the bubble deflates. This means dumping the risk on retail. There is only a finite amount of money available to buy AI stock, and the SpaceX IPO sucked a lot up, so it's a fight between Anthropic and OpenAI who is going to get what's left. The loser is left without a chair to stand on when the music stops.

One more anecdote: People at our company are very excited for integration of existing industry specific tools and processes.

Some of them do require frontier models but many can do with localish models. Some of these processes (in our industry) are also not really limited by software at all but by the speed heavy machinery can be operated and/or manufactured and/or transported.

I'm certain LLMs are going to continue rapid growth even there but it may be an ideal field for have lesser models work for longer if it saves money in the end.

Of course who knows in the end. I'm certain there is a high chance of this being entirely wrong

I am a software developer and I don't see a huge boost since many months.

Actually I see very varying results in my work.

Sometimes I can let an AI refactor an entire codebase and then (in the same codebase) it generates absolute shitcode.

because revenue growth is much more important at this stage, and investors are underwriting based on the first and second derivative

We talk about revenue far later in a company’s lifecycle than we did before because high-tech, high-capex companies take a long time to become profitable.

Amazon is now the posterchild for this, but it happens in other industries like pharma a lot too. And in today’s stock market, where a fundamentally weak company like Tesla can trade at eye-watering multiples because of speculation, why worry about profitability at all? (Partly joking but that’s how people think.)

When and why we start expecting Anthropic to be profitable are great questions though.

Completely anecdotal but I did find the better engineers tended to take to the AI tools quickest. At the beginning, tokens was probably a good way to pick them out. “Good” being entirely subjective around a product focus/mindset. If you ever find a good metric like that, it is critically important to never tell anyone, else it will quickly become not good.

From what I could tell, it probably even distracted from the metrics that traditionally matter. I seem to remember tokenmaxxing dashboards being built, and other "shovels."

> I don't think the "tokens consumed" metric was a very good one.

When mid or high level tech leadership pushes from top-down new things, that look dumb to the engineers, the leadership is either stupid or there are hidden incentives that regular people don't see. In my experience mid level usually matches with the former, and top level with the latter.

Because not only is there no guarantee that the peddled slop will be online or maintained or supported for any amount of time after its debut, but it's safer and more effective to just reap the slop as context into something slightly less sloppy.

Behold, the beauty of the modern internet.

For real. I'm so tired of all this, I hope the crash is swift and we can put it all behind us.

At the beginning of coding agents really picking up adoption (I guess 12-18 months ago now or so) I remember this sort of thing being so common that some startups were trying to sell a turnkey solution for "leaderboards".

I'm not aware of any of those products really sticking around.

We can finally dispel the myth that management is competent when they, almost as one across industries, tied continued employment to setting fire to company money

It's a combination of natural language being an awful programming language and frontier ai companies optimizing for longest time horizon at 50% success. I mean what did they expect? Length of time to not check in and still bumble around to turn a single prompt into a literal half-assed solution.

I don't have a few billion years to deal with stochastic evolution of my software, so I think I'll stick to intelligent design, thanks. Humans would be a lot better in a lot less time if they had been intelligently designed, too.

The very first goal post for AGI must be that it can act, learn, and grow independently. They must be self healing, self correcting, and anti fragile. That's definitely what the G in AGI is about. But all models fall into decoherence if you let them work on their own too long.

If you left an AI alone for a long time and when you came back they had built up a working society and learned new useful things, then that's the proof it's AGI. From this definition you only know it's AGI after you let it loose.

These AI companies changed the definition of AGI, to mean can do all white collar work.

I don't understand how anyone can call these AGI. I've always had the view that AGI == Independent human-level intelligence.

At some point someone moved the goalposts for this to include LLMs that emphatically are not that. I think it was Altman when they came out with Astra?

You just wait, we're still early®.

If you have to convince someone it's AGI, it's not AGI.

Or it secured more compute early at a lower price.

Also model/token efficiency.

List of possibilities is big

I know firefighters and machinists heavily using it, the only places I see slower adoption is purely because of bad leadership- exclusivity contracts with certain providers or just plain fear.

my brother uses it to do ocr on his books but, lately, claude has been refusing because of copyright or whatever so i've had to do it for him (using ai, yes, but not anthropic's)

I think this dynamic is underappreciated. Another thing I see in working with different companies (consulting) and discussing with colleagues at different jobs is that we as humans tend to do a pretty bad job of modeling business outcomes and the dynamics that govern them. There are lots of outcomes where you can identify one contributing role, increase the productivity for that role by a significant factor, and still ultimately not move the needle much, because the constraints span multiple functions.

As an example outside of software engineering, a lot of people in marketing roles have found agents to be incredibly useful for particular tasks. I've talked to multiple people who wouldn't standardly be classified as "technical hires" who are using Claude to manage things like digital advertising. It's valuable here because managing campaigns on a platform like Google or LinkedIn involves a ton of granular tweaking and analysis. Every performance marketer makes tradeoffs around what they're realistically able to spend time doing, and what makes the biggest impact. So if you have 250 campaigns that you manage solo, you likely don't have the time to really manually optimize the bidding and budget for each campaign every couple of days based on new data. You can only try so many new angles, because creating new media is time consuming. You don't have time to construct custom experiences on-page for every particular campaign. There's a whole industry of tools that turn a tidy profit automating parts of this for this reason. And agents are really good at it, because a lot of it is just applying the same analysis over and over again.

But for a lot of startups, increasing traffic from ads 20% while decreasing spend is a good outcome, but it doesn't actually turn into revenue. The product itself might just be limited, and already being discovered by the narrow segment of potential users who find it valuable enough to pay. Or the sales process might be kind of busted, meaning the team will really only close a certain volume of new deals in a given month, no matter how much inbound they have. Or some combination of all of the above.

If it's just one marketer using an agent to optimize their advertising, there's not a lot they can do about the other constraints. And so what you end up with is one employee who feels really productive and sings the praises of Claude, but not a meaningful lift in the higher level outcomes you care about, and the analysis gets messy. In an ideal world, everyone realizes where the other bottlenecks are, but that's a genuinely hard thing to do and people have complicating incentives.

I don't think this is a particularly new problem. It feels identical to any other situation where a department asks for a tool, loves it, and it doesn't really make a difference to the bottom line or make anyone else's life easier in noticeable ways.

> very capable chat-based UI that can automate a lot of tedious data entry

Because rewieving chat bots data entry is even more tedious then doing it yourself. And as long as they have to review it, it does not even matter that it makes only few mistakes.

I see it with JIRA and Rovo. It just sux using it, despite it being reasonably competent at creating tables and what not. If I could do markdown as used to be normal, it would be appreciated a lot more.

We are capped by testing and PR review. The junior devs can spam out a million PRs a day but we keep having production incidents from not carefully scrutinising the changes. So the speed of code generation is limited by how fast we can actually accept those diffs.

Right, this is what astonishes me about the constant rhetoric/hyperbole about how this is going to impact "all the jobs" and make everyone wildly more productive than ever before. It's a fabulous tool in a limited number of areas, and it's otherwise only handy here and there. But just because it's a fabulous tool for the coding that is so prevalent among HN users does not mean that it's a fabulous tool for everything.

A hammer is fantastic for driving nails, but it's not really all that good at washing a dog.

I use classic engineering to refer to the (engineering) disciplines where mother nature messes your day up rather than obscene levels of man made (or machine made) pure complexity ruining your day.

It's easy to understand how LLMs got so good at coding, they can rapidly iterate, self test, and it all happens in their digital universe.

These models now starting to spit grey beard knowledge of precision analog electronics, lightweight fastener design, and chemical surfactant systems is something very very recent. A few years ago I think we were pretty confident that the tech would hit a wall at the digital divide. Well, I can tell you now that it did not. It seems it kinda blasted right past it.

Vibes based analysis is very high volume right now due to upcoming IPOs.

On HN, can you please not be snarky? We're trying for something a bit different than internet mean here.

The idea is: if you have a substantive point, make it thoughtfully; if not, please don't comment until you do.

https://news.ycombinator.com/newsguidelines.html

I can hear the landfill full of those 3D glasses for home TVs shout in agreement

For big ones it’s often network effects and/or friction. Those can be pretty good moats even with LLMs. These models are also nowhere near the point where they can actually reverse engineer complex products.

Oh how badly I wish this was true.

How much easier my life would be if Chinese open models were even just 1 year behind outside of X posts and one-shots.

My easy litmus test is would the year old model (Opus 4.5) being open sourced today would eat the traffic of every single Chinese model... and the answer is pretty much yes.

Every Chinese model would throw away their work just to post-train or distill Opus 4.5 and get a more reliable model.

They can't keep getting away with it.

Amazon hosts models. And?

Why do you say this? If they get to even 100B RRR they will pay off all of their commitments in <10 years with money to spare. Does that seem so impossible to you?

It's also very damaging to under-value, I think OpenAI got hit hard with lack of compute and is still struggling.

He says a lot of things, but one specific thing I remember from weeks ago is that he said Anthropic growth slowed on Q3, and that's why they decided to push for the IPO (and use only Q2 numbers).

Please don't post swipes about people on HN, no matter who it is or how right you are or how right you feel you are. https://news.ycombinator.com/newsguidelines.html

The thing is I'm not spinning up a new app every day, I'm just working on a large existing system with enough constraints. From that perspective tools + the general rigor I already try to apply means that I can get some value out of LLMs without waiting for it to do something "good".

"I can one-shot a new thing" is interesting but definitionally a thing is new only once. Maybe if I worked in a space where we had way more throwaway things to work on.

At the end of the day I want to understand the output anyways, so it one shotting a thing vs me going back and forth with it in smaller increments is a mixed bag anyways.

You could in theory be responsible for the destruction of the entire solar system in order to pay you. Sit on your throne atop a mountain of skulls and weep.

Their growth was positive. The second derivative (growth of their growth) was likely negative, and that's what this article is about.

Growth is first derivative. Deceleration is second derivative. Focusing on second not first I think was the point, not implying there is some nth derivative under discussion.

Winning an argument online 101: call your opponent desperate.

Question one is how do their investors get paid back. I found Series E through Series H totaling $111.5 billion in investments. Since series E was $3.5 I’m assuming all the previous rounds added up to less than that, so let’s say 115 billion at the outside.

That debt will transfer to public shareholders at the IPO but also will double or triple in the process, as the investors expect to be paid very, very well. So the current shareholders are looking to extract something over $400 billion from the IPO.

With the current revenue they might pull off that P/E. But they’ll still be extracting some of that value from the customer base, because shareholders will demand it.

The contracts might, but do they pink promise?

The company was founded under the principle they will violate laws to acquire training material

The subscriptions do not require training on your data.

Ah the beauty of circular financing. By owning the company who pays me to provide a service, I stand to gain massively by giving them a discount, making them look profitable, and fooling investors in to providing my exit liquidity.

With AI who can tell.

I am just imagining Anthropic (and OpenAI) employees reading these comments and laughing at the rank ignorance of the commentariat.

I didn't think I needed to spell it out for the HN crowd, since folks talk about the Chinese models every day. But, OK...

At least three Chinese companies have bragged about their latest models being trained and/or served on Chinese-made GPUs. Literally within the past 24 hours, LongCat spent a large portion of their release announcement for 2.5 bragging about how efficient their training process is and that it runs on domestic (Chinese) GPUs: "LongCat-2.5 uses long-sequence training during both mid-training and SFT. With less than 80 GB of memory per domestic AI accelerator, the training system prioritizes memory efficiency and aligns its parallelism with the supernode architecture."

https://longcat.ai/blog/longcat-2.5/

DeepSeek has repeatedly blown the curve in terms of how efficiently a model can be trained and served, Xiaomi and Z.ai aren't far behind, and have also discussed training and serving on home-grown hardware. I'm sure you can search Arxiv for all of the papers from these providers (DeepSeek, in particular, who document seemingly everything, even when they're doing something ahead of the US providers) as well as I can.