criley2

Considering that the biggest and most successful play for SpaceXAI in the AI space has been "Sell compute to Anthropic" I also wouldn't say that all of the grok junk has been obviously successful, either.

The way I think of it is that "20x" means "You get up to 20 tokens for each 1 if you run your limits down".

For example last week I put $1400 thru my Anthropic enterprise $200 plan. That's $5600 per month, or 28X the $200 my company pays (according to `npx ccusage`)

There is a whole side of sci-fi that has explored the idea that AI government could finally cure us so much more than just one class of corruption. There are, however, new downsides.

I use Opus 5.5 at work and I use Sonnet 5.5 at home. Sonnet 5.5 medium according to AA cost per task is $0.48 per task, while GLM 5.3 (max) is $2.01 (opus 5.5 high is $1.82). GLM 5.3 flash is $0.28 per task and Deepseek 4.1 flash 1017 is $0.26.

So it seems like Sonnet 5.5 medium, GLM 5.3 flash and Deepseek V4.1 flash 1017 are all in a similar space in capacity and price, however the $20 anthropic sub will get you like $200-$300 in sonnet 5.5...

Another bonus is that sonnet uses a fraction of the overall tokens and finishes much faster than the average chinese model, so your work goes faster (Sonnet 5.5 medium in 2 minutes, Deepseek 1017 in 6.5 minutes, GLM 5.3 flash in 13 minutes...)

I still don't get why paying for chinese inference makes sense. I should try Haiku 5.5 high/max because it's posting higher numbers than Deepseek V4.1 flash for half the price

That senile President didn't order a violent coup and call for violence on social media, therefore he didn't earn a ban.

He also didn't rape children and steal billions while destroying the world economy...

Hey, if you want to argue with the parameters about what stereotypes come out most strongly for that prompt, be my guest. But no need to shoot the messenger. LLMs are basically stereotype machines, it just turns out if that you make a trillion stereotypes, you get a lot right.

Honestly, my first impression is that the model believes that "software for a US persona" is less secure, and potentially more <something else> that wasn't tested.

Sounds about right. As we can see with the US approach to AI, it really is about maximizing the investment as fast as possible, maximizing the utility as fast as possible, and blowing past any concept of safety or responsibility.

I mean these models will have at times radically different output when the only difference between prompts is that one prompt is lowercased.

I'd love to see the reasoning traces because if you write "software for a US persona", it will try to fill in the blanks. What did the user mean? "US persona"? Prioritizing profits, advertising, and shipping new features over safety and responsibility? That's what many humans would assume you meant...

An obvious downside to plan mode is that without access to the edit tool, the model can't confirm its hallucinations.

Removing edit tool from your model is like lobotomizing your engineer.

Here's the secret: Just building the thing is the modern "plan mode". You don't have to keep what it made. But a completed, verified, and tested output is always more valuable to my planning process than a hallucinated sloppy guess.

No "may be" about it. There is no benefit to having bad models write your code, and there's no benefit to "plan" mode. Plan mode is an outdated 2025 concept that should be removed from standard harnesses.

There are 2 scenarios for dumb builder agents:

1. Your "planner" actually implements the project almost completely, finds all the gotchas, solves all of the emergent problems, and fully plans the solution.

2. Your "planner" takes a half-assed first attempt, solving none of the major problems that will arise during implementation

In scenario 1 -- just pay cache output on a few more tokens to finish the project. No need to spend twenty minutes pumping a few million tokens through deepseek regardless of how cheap it is to do so.

In scenario 2 -- good luck getting quality output

I do think this "software team" LARPing is an engineering version of AI psychosis. You can get all of the little parts humming along, sending messages, and looking really cool. You can even convince yourself that all of these busy little bees actually make better product.

But a frontier model working alone can produce better output in a tiny fraction of the time. Just build the thing.

I certainly don't think the self-reported single benchmark scores are the best way to answer the question. I waited for the Artifical Analysis run (which itself is still mediocre, but at least it's an order of magnitude better), and predictably the results tell a different story.

Sol 6.1 High: 50 intelligence, $0.32 per task, 4.4 min per task, 13k output tokens per task

Sol 6.1 Medium: 48 intelligence, $0.21 per task, 2.9 min per task, 8k output tokens per task

Deepseek 4.1 1017: 45 intelligence, $0.26 per task, 4.8 min per task, 89k output tokens per task

If you look at the actual pareto frontier across the full benchmark, Sol 6.1 remains firmly on the pareto line, while deepseek remains firmly under it, just as it was before https://i.imgur.com/N72p5NF.png

This result is exactly as I expected, and seems to run counter to the effusive praise of this thread. Sol 6.1 Medium is producing higher quality output at a lower price with nearly 10X less output tokens! And this is not surprising.

And finally, Mimo V2.6 Pro continues to look like a stunning release, clearly smarter, cheaper and faster than deepseek v4.1 1017

What would you prefer, 30t/s 50k output tokens needed for solution, or 300t/s, 500k output tokens needed for solution?

This is why "tokens per second" alone is a deeply flawed comparator.

This is also why discussions of Chinese models focus on tokens per second- they struggle to compete on intelligence so they dramatically increase effort levels to compensate, meaning they require massive total output tokens.

For example, flash 1017 isn't benchmarked yet, but 0731 required 15.4M tokens to complete the artificial analysis benchmark, and Sol 6.1 high required just 2.8M.

Sol 6.1 uses 2.8M at 64 tok/s = 12h 9m

Flash 0731 reported 149 token/s at 15.4M needed = 28h 42m

And in my example in the first line, scenario A (30 t/s) completes 6 minutes faster than scenario B, despite 10X slower inference.

Remember, the vast majority of pivotal Phase 3 trials are funded by for-profit companies.

Twenty years ago, Rick Ng's Drugs: From Discovery to Approval cited benchmarks of roughly $1 billion, 10 years, and 10,000 candidates to get a single drug across the line. Today, estimates easily run $2 billion to $4 billion.

Given that level of financial pressure, especially in Phase 3 where a single failure can wipe out a biotech company, assuming humans "follow the spirit of the trial process" is wishful thinking. Humans already exploit the system through p-hacking, shifting surrogate endpoints, and cherry-picking subgroups. An AI gaming trial parameters wouldn't be a new problem. It would just be an automated version of what sponsors already do.

In this case, I was using Hugging Face inference set to one of a few American providers. For my professional work, I do not use Chinese inference.

To be clear, I'm discussing non-technical use of Gemini and ChatGPT through the web harness.

While I do rarely use the web harnesses for technical research, nearly all of my technical questions are going through coding harnesses and layers of my local context, skills, etc, so what I produce from codex/claude code is not comparable to what others produce.

So my comparison of the web harnesses is me asking questions about caring for my baby, or product comparisons of new televisions, or visualizing a room organized differently (google sucked at this so bad, chatgpt did great), or asking for help in a video game, or other normie AI uses lol

I will say, Gemini is the best for asking about google services, such as Flights or Youtube.

Personally I find this shocking. I don't think our application is that complicated, (typescript full stack graphql reactnative etc) but deepseek 4.1 flash is a bumbling fool, junior-level at best, who takes a very long time to make a very big mess. Opus 5.5 one shots truly impressive code in 5 minutes, while deepseek 4.1 flash takes 20 minutes to do horribly. I simply don't understand how folks claim they get good engineering out of it. Maybe we still care enough about the fundamentals to notice the mess ...

I feel like whoever wrote this doesn't use these models regularly. Deepseek v4.1 Flash is far from the pareto line. You can get the same performance for half the cost from Luna or Haiku 5.5 now, or you can get substantially improved performance at the same price with Sol 6.1 ~medium.

It did correctly make waves when it launched, but was quickly eclipsed by the deluge of american model releases, especially those competing on cost.

In my experience

- Gemini 3.8 has a beautiful answer. Typography, use of lists, brevity, style, even the font choice on the web harness -- all of it is A tier or even S tier. Notice I didn't say answer quality. The answer is always extremely mid, and the harder the question (or more effort needed to answer it), Google just quits. I give it a "C" on answer quality

- ChatGPT (Chat, pre 6, I assume 5.6 although they commonly hide the model picker): Ugly answer. Runs on printing page after page of unnecessary side notes. Mostly just walls of texts in paragraphs with little thought to information design. However, inside of that mess of text is almost always the answer I'm looking for, and it's almost always incredibly better than the truncated, desgined Gemini answer.

For a while I ran all of my question queries on Gemini and ChatGPT, but I noticed that I basically always picked ChatGPT's answer head to head.

>"Minimize ambiguity" has been my go-to instruction as of late when the agent drifts back towards vague terms and lack of specificity.

Anthropic has called the greater category containing this type of writing "mannered prose" https://platform.claude.com/docs/en/build-with-claude/prompt...

If you ask the models to avoid mannered prose (or use their extended prompt), it basically eliminates all of this type of slop writing.

Here's a de-slopped example.

> Write Like It's 1866: LLMs Relearn Telegraphese

> Adding one sentence to a prompt, telling the model to write like a telegram, cut its output tokens by 40–49%. The sentence asks it to drop articles and filler but keep every fact. Models from four different labs then answered questions from that compressed text as accurately as from normal English. So when one model writes something for another model to read, you pay about half as much for the output. This post introduces the Telegraph Test, a benchmark that measures how well a given model does this.

I don't believe that "ethical" and "ad supported" are compatible. Harvesting our private data and selling it to the tech-dystopia to further curate profiles about our every move can never be considered "ethical", imo.

From my perspective, those people who are taking this public client side code (not emulating any kind of server), and removing the privacy nightmare, are actually doing good for society. The software is more usable, more performant, and far more secure when they are done. The only harm is the authors ability to monetize.

I don't think it's possible to have a fully client-side web product and be able to enforce strict guardrails on the use of the code. Regardless of ethics, it's just not feasible. What you give up by delivering the full source code to the browser to render is control over the source code.

If the author wants more control over their source code, and easier monetization, they should compile a binary and distribute that. The guardrails protecting source code, duplication, and copyright infringement are much more clear. That's just the harsh reality of delivering source code to clients.

America's geriatric lawmakers don't even use email. They're decades away from understanding AI. Any laws in America will be written by the industry itself. Generally speaking, that means regulatory capture and the entrenched big players shutting the door on any competition. Anthropic will help us get safety laws that, surprise surprise, only Anthropic models satisfy. And all those pesky Chinese models will definitely be banned first.

I have a totally different read on this. I also play a bit of minecraft many days, and for maybe 15-30 minutes at a time. Mike Tomlin here was playing on creative -- no risk, no survival, just infinite blocks to curate his zen garden. This is the perfect way to wind down for 15-20 minutes after a tough day in a hotel room. Or where ever. What percentage of NFL coaches or players do the exact same thing except with a phone instead of a controller, do some doomscrolling for 15-30 minutes and move on? This feels like just a different flavor of the same type of retreat that many people do.

Deepseek Flash v4.1 is only "40X cheaper" if you do not account for the time of the engineer reading the output. If Opus 5.5 high requires 1/2 of the actual engineer time, and the engineer costs $100-$200/hr, then Deepseek v4.1 is actually the more expensive model to use.

I have tried your workflow many times, and simply letting Opus do the implementation costs much less than wasting hundreds of millions of tokens letting deepseek and opus go back and forth and back and forth. And bonus, my project finishes in 5 minutes instead of 20.

In my experience it never works well on any real work. In fact, I'd go the opposite, plan with the dumb model and execute with the smart model because at least the model writing the code and solving the emergent problems is capable.

In my experience (and I've been trying this a bunch): smart planner + dumb executor produces worse code with higher spend than simply using the smart planner to do both.

It's easy to understand why:

- If the planner has truly thought the issue through, properly designed the solution, solved all of the emergent problems, then the final "write" of the code is just a few more output tokens.

- If the planner has NOT truly planned the issue completely, then you're letting a substantially dumber and less capable model make significant decisions, and trusting its problem solving, without having a better model check it.

If you're highly cost conscious (paying for your own tokens and not making any money) then you have no choice but to trade your time and effort for tricks like this to save money by lowering the quality of your output.

But if your employer is paying for tokens: just use the smarter model. You save your time preventing re-work and reducing code review, you save your employer money (primarily from the cost of your own labor and reduced rework), and you get a better output every time (Opus 5.5 mogs Deepseek 4.1 flash in every single way except cost).

Conflating "the internet" with "~Trillionaire owned, oligarchy controlled, American social media" is certainly a choice.

I'll say the opposite: Every single place that has banned American social media is better for it. Every group of people denied access to American social media are healthier and happier for it.

So if you bought a bottle of water, and inside of that water was a poison that kills you, intentionally added by a criminal because it makes them more profit, that's not murder?

Those people were doing a line of coke, they had no idea they were about to overdose on an opioid.

Tiny bits of fentanyl are added to a variety of recreational drugs used extensively by all kinds of people, rich and poor. They add it to cocaine, heroin, meth and even marijuana. They imitate oxy, percocet, vicodin, xanax and adderal using fent laced pills as well.

As much as 20% of Americans use these recreational drugs. Europeans on average use recreational drugs like half as much as Americans, maybe 10% use them.

Fentanyl is made by Mexican cartels from Chinese precursors. I suppose there aren't many Mexico's with massive organized crime networks in Europe to pour deadly drugs into their countries.

No one "does" fentanyl, it's a highly addictive additive that is secretly added to other things. Even so, 50,000 - 70,000 Americans per year die to it (more than suicide and gun violence many other things).

This is an "American" thing because fent primarily comes from Mexican cartels (Sinaloa and Jalisco New Generation).

They buy precursors from China, they produce it in labs in Mexico, and then smuggle it into America where they kill 50,000 to 70,000 Americans per year.

Not getting political here -- but what would you think if organized crime in a neighboring country was murdering 50,000 to 70,000 of your countrymen per year?

Developer comms are very important. I've often "fixed" something that was working exactly as the designer made it. It certainly wasn't a bug in my work though. I've worked in places where we call this a "design omission" ;)

They're not fixing it, they're "changing" it, which makes sense, because the design changed.