Harness improvements like context pruning, malformed tool call healing also very helpful to boost smaller local models. Doing alot of work in this area lately.
Qwen 27B in Terminal Bench 2.1 73->85 for instance.
Doesn't matter if the product is real.
The idea is to cultivate a sense of wonder and adventure.
To get people talking about you.
Then you sell more of the products that are real.
Doesn't matter if the product is real. The idea is to cultivate a sense of hope that the $250,000 deposit should stay where it's at, no matter what people talking about me think of that idea. At least I haven't bought a cyber truck.
Real?! From these circus folks? Their (lack of) sense for reality (practicality, safety) you see in this circus is scary! Same folks and mentality make those other ones too that you are supposed to use!
I am sure someone did it on the back of an envelope somewhere. It probably wouldn't translate well to the general public. Probably best to avoid such things
in the 1990s i read that ai wouldn't be competitive with biological brains in real time because the neuron density of the latter was so high that heat generated by the silicon would burn itself up. perhaps it is still true as llm are frozen at a fixed time, so it avoids real time.
In the very long term, no. Can't beat thermodynamics. But ideally I'd prefer if life survived the next decade at least. That's looking unlikely with current AI progress.
That's not their business model. They won't win on price, but they won't compete on price. Their business model is making the current state of the art.
If I'm a business and I need something done today, and bc Anthropic has the best model, there's a 99.9 chance it will be completed successfully for $1000. And using Deepseek there's a 70% chance it will, for $10 - you or me will go for the $10. Big businesses don't. Bc 1000 per task is nothing to them.
Actually opposite occurs. Big businesses are ok with a mediocre but cheaper result. Very few are willing to pay such cost. Just look at tech wages and the distributions
The business I work for is absolutely sensitive to 10 vs 1000, depending on the task. And it's a multi-billion dollar business. 1000/task may not be much on it's own, but there are a lot of tasks.
Big business doesn't pay more for better, but the do pay more for predictability, support and targeted outcomes. They will happily trade a chance at 100% better results for 10% less chance of unplanned outcomes
Except businesses are going the opposite direction here. The lack of stickiness makes the “premium” argument hard to play. Oracle won because swapping databases is a giant PITA. Swapping models requires almost no effort for most uses. And because of that enterprises are all building model marketplaces where providers have to compete on price performance.
Most folks I know can choose from any of the big labs or open weight models and they get billed internally for tokens against their budget. There’s little incentive to no switch to the lower cost closers.
This setup is a nightmare scenario for the big labs trying to execute the traditional enterprise sales plays. Those only work if your product is sticky and AI models are one of the least sticky things in the history of tech.
Cosmo Kramer: It's a write-off for them.
Jerry: How is it a write-off?
Cosmo Kramer: They just write it off.
Jerry: Write it off what?
Cosmo Kramer: Jerry, all these big companies, they write off everything.
Jerry: You don't even know what a write-off is.
Cosmo Kramer: Do you?
Jerry: No, I don't.
Cosmo Kramer: But they do, and they're the ones writing it off.
Edit to be fair: the "500 minus tax write off" comment does make sense but often people overstates the benefits of writing things off.
My take - the $500/month is likely similar to the $200/month plan (albeit with Astra Ultrafast that consumes tokens ultrafast).
Or maybe it gives you just a little bit more tokens than older $200/month plan for plausible deniability:
Maybe me: “you’re charging $500/month for what was $200/month earlier”
Maybe OAI: “no we’re not!! You get *more* tokens than the older $200/month plan and did you hear, you get Astra Ultrafast access!”
Maybe me: “oh really? How many more tokens??”
Maybe OAI: “I’m sorry we can no longer share that information, it’s confidential. National security and all…”
More than an hour after I wrote this, OAI sent an email sounding eerily close to my approximation:
- $500 plan comes with not 20x,
but 25x allowance!
- also Astro Ultraspeed
- $200 pulled down to 10x quota
Things I hadn’t anticipated:
- one time ~62,000 credit
top-up ($2,500 API call
value)
- 5 hour quota limits removed
The removal of 5 hourly quotas can negate some reduction in token allowance. Especially relevant for those of us working in 2-3 day bursts (such as weekends). Because these folks weren’t using their full weekly quota before, they already had a decent buffer.
Maybe OAI thinks people are creating new unnecessary work to use their allowance completely. Well they aren’t entirely wrong. But the experimentation has been useful in letting us discover new use-cases, and to develop and flex our agentic muscles.
I've talked to VP at my company and we both agreed $1000 for the time saved, is still 'worth it'.
Dev salaries are high and devs often have to realign buttons or change colors. We are not always working on complex ideas.
I rewrote a big piece of software from scratch....it had 0 docs. Documenting of classes and flow took me weeks. Complete rewrite took almost 2 years. I am confident that with AI I would be able to get this done in a few months.
I mean, it's a good deal for the boss, but I also wouldn't want to waste two years of my life doing something I could do in a couple months or weeks, would you? That's not a good use of my life, regardless of whether someone's willing to pay me.
I'd rather use the fancy models while absurd valuations are subsidising them and then pick up a couple grand worth of their old hardware in a few years when they're eol.
Not 100% in the first year (where I live). By the time it’s depreciated, it’s time to replace. Subscriptions are 100% deductible in the year you pay them.
Wouldn't work. A dynamic but verifiable problem. Is just a perfect target for a RL environment. If you don't have the verifiable part the benchmark is useless, or really expensive with human review. (Or just open ended)
There doesn't need to be a single correct response. Generate "novel" problems with a set of acceptable solutions/outcomes, then verify that the response satisfies the criteria.
The core idiom I've used for years is deliberately tiny: fully encapsulated custom Web Components that simply re-render when matching an attribute and `window.state` changes. Data model is simply:
window.state = new Proxy({}, {
set(target, key, value) {
target[key] = value
document.querySelectorAll(`[data="${key}"]`)
.forEach(el => el.render?.())
return true
}
})
Got off Gemini before they ended the cli, was not working for me, and went to Codex and opencode/deepseek flash v4 0731 which was working ok.
But got the google ai pro free 1 yr student - been liking agy quite alot actually - pretty good weekly allotment. Like the default, Flash 3.7 high.
Even bought extra credits I've been hammering it all day long.
People talk all day about coding but where this thing amazes me is devops. It hooked itself to discord so i can command it, set up or troubleshoot ssh servers on my home network. Blew my mind a couple times.
reply