criley2

It's not free. You're paying electricity and you're ignoring the cost of the hardware. Even on electricity alone, there are cloud providers who may beat your laptop on price per million tokens. Qwen 3.8 flash is interesting in this space.

Not to say that there aren't other benefits of running models locally, I loaded Qwen 3.8 27B 6bit MLX just yesterday.

Sonnet 5 is the worst model of 2026. Literally just turn effort slider down on Opus, it's smarter, faster and cheaper than whatever Sonnet is.

Beyond that, I find this whole plan and build thing to be a pointless waste of tokens. If your planner made a detailed enough plan, then the cost of executing that plan is a just one turn more of cached tokens, and minimal time.

Meanwhile: switching agents, reloading context and building from the plan will easily balloon your token use and time. And any emergent problem that the dumb executor finds will instantly wreck the implementation because they're not competent at solving it. And if your plan is so perfect that there's no edge case then you're wasting tokens because your planner was one turn away from finishing the project via cached tokens.

Those prices are just tokens? Since each model uses different amounts of tokens to do the same thing, it's a misleading price that often makes open-weights look more competitive than they are, since most open weights models use dramatically more tokens and time to complete tasks than many frontier models.

In Artifical Analysis's cost per task, Luna(max) costs $0.05 per task, and Qwen 3.8 27B costs $0.25 per task, a 5X increase. We'll see how 3.8-flash-next does.

I absolutely experienced this in college. I signed up as a computer science student, as one does. I took all of the freshmen classes across broad topics, and the first biology class was basically just like the article describes. Words and memorization and microbiology and boring stuff. Who cares, let's get back to the dorm room to play CounterStrike on the school LAN.

But then I took the second required biology course. And rather than be about memorizing the krebs cycle, it was macrobiology. Ecology, evolution, and everything I hadn't realized I wanted to learn about the world around me. I instantly fell in love and 99'd the class. I changed my major to Biology, went pretty heavy into biology and chemistry, and didn't take more than a few Java classes in Comp Sci. I graduated with a B.S. in Biology.

I'm still a software developer, mind you, but I wouldn't trade the Biology degree for anything. Learning statistics, experimental design and experimentation, and just having the opportunity to deep dive into the physical world from the biggest to the smallest details, it was life-changing.

I still ended up learning data models, algorithms, and probably would be a better engineer if I had dug in deep on comp sci, but I like to think the biology degree helped me be a better developer for different reasons.

I append this to many of my opus claude code prompts

`You may use a Fable subagent to answer questions, solve problems, and provide an adversarial review of your ideas and code`

You can use a similar pattern in most any harness, and you can tell them to use other harnesses. In claude you can write `Use codex cli to have Sol56 Xhigh provide an adversarial review to your plan before presenting it to me` or `Use opencode cli with GLM 5.3 to verify all code review findings before presenting` or whatever you're doing, as long as those other tools are setup and ready to be called.

IMO: This isn't useful as a token saving pattern in my experience with agentic engineering, but it is useful as a quality-enhancer.

It's pretty easy for the US to functionally ban chinese models. They only have to target US firms like inference providers or the biggest users, and pretty much the whole domestic market will fall into line. They don't actually care about the final few %.

Regardless of whether or not adversaries are using them, the US has by far the most compute available, and we've now hit the line where major providers are no longer releasing their best models. The public gets the "current" level of intelligence, while the US government gets to control access to the actual frontier of non-public AI. From their perspective, their enemies using GLM5.3 while they have GPT6 and Mythos6 or whatever is a fine trade.

I don't support a ban at all, nor the US's behavior, I'm just pointing out some facts that change the argument.

I totally agree - designing a competent AI agent with a fully customized harness to successfully pull off this task is a much more challenging engineering effort than merely creating an ordinary computer program. Had OP made chatgpt write an ordinary program instead, they likely would have succeeded in their task.

What a sloppy reply. You've hijacked a thread on mathematics first to complain that your incompetent attempt to use ChatGPT to find a job failed, but it seems now that this was a ruse to instead begin arguments unrelated to the article at all where you just spam arxiv links you've never read to "prove" that AI is a scam.

This comes across, frankly, as either Dunning-Kruger (classic illusory superiority), or potentially as mental illness. The slop dump is highly reminiscent of how a schizophrenic friend of mine communicates.

Do you really think slopping down a bunch of random arxiv links "proves" that AI is a scam and you're so smart and everyone else isn't?

Most awkwardly for your arxiv slop -- most of this is irrelevant to your central claim, and you've missed papers that are much closer.

For example your LogicGraph paper: "Can't exhaustively enumerate all minimal proofs" is not "can't distinguish Ireland from London".

Or your "Do VLMs Understand 3D Scenes..." is nothing more than citation decoration, completely irrelevant to our discussion.

Or your "Frontier LLMs Still Struggle with Simple Reasoning Tasks" which is potentially your pièce de résistance, it supports brittle multi-step constraint handling, but isn't remotely an eval of a modern web-search agent.

For example, VibeSearchBench would have been far more relevant to your claims https://arxiv.org/html/2605.27882v1 (but still obviously not proof that AI is "a parlour trick")

Going further: my point that we need to discuss your beginner's approach to the harness is substantiated clearly here: https://arxiv.org/html/2605.23950v1

Finally, failure to exhibit human-like generality is not evidence of absence of intelligence. It is evidence that whatever cognitive machinery LLMs possess has a very different error distribution from ours. Your General365, LLMEval-Logic and the Reversal Curse are actually fascinating evidence for that jaggedness, rather than proof of your claim that AI is a scam.

A business does need a small number of their most senior engineers doing high altitude work that can, at times, include helping sales estimate new features. But in my experience, it's not rocket science and a good product team can do this on their own with a quick async check over chat. At most, a single meeting is all it takes.

I've heard of Sales Engineers as well, embedding programmers directly with sales teams.

But the vast majority of programmers should not be spending any significant amount of their time on this. Their value is in building and scaling well-specified systems.

I think you're confusing product and engineering. I get that programmers are smart so we just assume we can do every job, but it's a waste of your time and salary to talk extensively to customers and create product requirements. Let the PM's run the user research sessions, you can find more productive things to do.

However, where we agree, is that there is more to engineering than writing code. It's problem solving. Even if the product team, the c-suite, the board, the investors, et al, are all in on a product that they believe customers want, doesn't mean the real problem of bringing that idea to life at scale has been solved.

Scenario one: you use software to connect to their server and download a webpage. You are a user.

Scenario two: you use software to connect to their server and download a webpage. You are a "bot".

Make it make sense

GPT5.6Sol completes the suite in 70M tokens, while Qwen3.8Max needs like 145M tokens. So this is a case where models like Qwen 3.8 and Kimi K3 use a lot more output (reasoning) tokens, go a good bit slower, so they can ultimately achieve a better intelligence score than if they went more quickly.

There are a couple of frontiers (ok bad word, maybe categories) in open weight models.

These Qwen 3.8 and Kimi K3 style models aren't trying to win on price, they're trying to compete on intelligence and capability.

Models like Deepseek V4 Flash (updated this week) are $0.03 a task, or 50X cheaper than Qwen3.8/Kimi K3, and 100X cheaper than Fable, while offering stunning intelligence. That's a different frontier for competition, and perhaps one more interesting for someone who wants to see them compete on cost.

While I agree with the premise that there are Thinkers and Shippers, I reject labeling of tinkerers and entrepreneurial. There's nothing entrepreneurial about working for a big business and shipping cool things.

But you're ultimately right, some people want to solve sudokus and some people want to change the world, and the sudoku solvers are really upset that a Sudoku Solver and World Changer 9000 device is widely available, because it spoils the fun they have with sudoku.

There is already something on HuggingFace at the level of Mythos. It's called Kimi K3 and it's running laps around Opus5, Fable5, and Sol56 at cybersecurity.

It's so good that the US government is rushing to ban all Chinese models as fast as they can.

>No, you don't. Without training cost you can infer only the marginal cost of serving this kind of models.

Are you talking about Kimi's training cost or the training cost of the model(s) that Kimi distilled?

Because Moonshot didn't even incur the majority of the training costs either

I don't think the invention of writing is as awe-inspiring as presented or as difficult/impossible for an LLM to achieve as is commonly believed. The invention of writing was a long series of micro-improvements over common every day things. Let's keep track of count by writing a line in the sand. Let's cut it into a tree. Let's push the line into clay. Let's use two lines touching. These are the kinds of things an LLM can also reason through and achieve.

I reject this idea that an LLM couldn't come up with something novel that isn't in the training data. We know it can come up with sentences that aren't in there. We know it can come up with mathematic proofs that aren't in there. So could it feasibly create something as fundamental as writing or language? Would we even be able to understand what it had done, or would it be so foreign to us that we wouldn't even realize?

Those early writers did have a training set of human knowledge that led them to pushing stones into clay. And our agents may very well have a set of knowledge that leads it to pushing their proverbial stone into their proverbial clay too.

Zen is nice, but they require US hosting so they don't get new Chinese models right away. There is no Kimi K3.

Go is nice for the ten minutes you can use it until your hit your cap.

No, and the reason is simple: Usage is bursty and if you don't maximize usage of the hardware you're going to lose on price.

Ok you can host this model once. What if I want a dozen subagents? Ok you can host it 12 times at once. What if we go a whole week only using max 4 at a time? Etc etc. The limits imposed by self-hosting might be bearable for a variety of reasons, but it's going to be more expensive and less convenient/useful.

> "Not a scam IMO"

> "I think SpaceX is a solid well run company"

This should be all the proof you need that it's a scam. Here's why:

The company isn't "SpaceX" it's "SpaceXAI", and nearly all of the valuation comes from the "AI" component, not the "SpaceX" component.

Totally disagree that the value of a project is it's durability. The value of the project is almost entirely disconnected from durability. Value is simply the ability to solve a problem for you. I'll use a bad screwdriver before I use my fingers, even if I prefer good screwdrivers.

Claude Code was a vibe coded experiment, a "what if", that basically consumed software engineering in six months flat. Not because it's durable (it's not), but because the value it provided was so overwhelming.

People will use bad software if the value it provides is high. People will avoid the most durable and battle-tested software ever written if it doesn't actually provide value (solve a problem for them).

在英语中，我們会说“Chinese”。

See, this is part of the confusion. There is no such thing as "GPT-5.5-codex". The last codex-branded model was "GPT-5.3-codex". Starting with "GPT-5.4" the main model handles agentic engineering and they did not release a coding model.

Both the web harness and codex app/cli use "GPT-5.5".

I'm struggling as well to understand, and I think perhaps they mean they use ChatGPT website with GPT-5.5+reasoning for problem solving, and paste the output into Codex CLI/App. I think they're saying that letting Codex CLI/App problem solve with GPT-5.5 isn't as effective. Essentially that the web harness is superior to the agentic engineering harness for problem solving?

Not sure if I agree, but I do happen to use a fair bit of web harness as well, just because I find it to be much more effective at web search and a different type of reasoning. So I must agree a little or else I wouldn't do that.

What a weird thing to suggest. This model is worse than some Chinese models from last year, which were already worse than frontier models from last year.

This looks like a worse Deepseek V4 pro. Costs more, dumber, slower and more verbose than most modern Chinese releases.

Seeing how those Chinese releases are in various states of open source, I would hope a competitor could at least match them.

Really cool experiment. A lot of jank. It would sometimes rubber band me back, movement was grid aligned in a way that made accessing the secret room challenging, and the whole tab unexpectedly crashed with no error. 5 star would play again

It's not fun to spend an hour going zig zag against the wind? What about an hour doing it with a boat full of precious materials that can't go through the portal while you're being chased by some ungodly horror from beneath the waves?

Valheim was a blast.

In your example, those projects were not the domain of engineering. Your friends couldn't afford expertise and those projects never would have been made before. Therefore, these are net-new projects that replace nothing that came before.

This is common in technology, especially software, that improvements in efficiency make software cheaper and expands the total pool of possible software.

And, even if one of the "one button shipping" platforms (and there are many) was hooked into the agent, your friends will hit major problems that they and their AI cannot solve. Whether its tech-debt hell, a major security breach, or something else, any sufficiently complex project will require expertise. Will be a fun day when a European regulator asks a friend about their GDPR compliance and their agent is like "shrug"

This is a non-answer, and in fact, a statement like "don't let fear be an excuse to lose rigor in thinking" in response to my question "how verifiable are their claims" is insulting and sloppy. Rigor in thinking includes human discussion and humans asking questions, but yet you shot that down.

ChatGPT, do what this user wouldn't, and answer the dang question:

> No, Apple cannot verify that Private Cloud Compute is completely immune to nation-state actors, contains no zero-days, or could never be subjected to secret legal compulsion. Nobody can honestly establish those absolutes for a complicated, evolving computer system operating across multiple jurisdictions.

> What Apple has done is more meaningful than ordinary corporate “due diligence,” however. PCC is specifically engineered to make clandestine access—whether by hackers, insiders, or governments—technically difficult, difficult to target, and more likely to leave externally detectable evidence...

> Against ordinary attackers, rogue employees, conventional cloud administrators and routine government data requests, PCC appears exceptionally strong for a cloud AI service.

> Against a targeted nation-state willing to combine zero-days, supply-chain compromise, endpoint exploitation, legal pressure and secrecy, the right description is: Highly resistant, deliberately difficult to target, and unusually auditable—but not immune.

Thanks ChatGPT. Don't know why I bother to ask humans anymore, it's StackOverflow the whole way down.

What does verify mean?

Can they verify the private cloud is completely immune to nationstate actors, has no zero-day vulnerabilities, is completely bulletproof in a court of law and can never be compelled to secretly share info with government(s), etc?

I think the users fear here is real. "We did good due diligence at the consumer level" and "we're completely immune to nationstate hackers and clandestine legal cases" are very different things.

You can do basically the same thing as cloudflare except as a skill you run in your local harness. If you're going through the motions with PRs and are familiar with actions, you can have it run in a github action instead. But this is basically just a skill. The Claude code review skill is a simple version of exactly this.