Post

The odds of 1000 smart people writing the code-equivalent of a nuke is smaller than the odds of a hundred million A.I.-assisted idiots writing it.

the issue is that as we move from human vetted code to vibe coding there is immense trust that the machine will not insert anything unwanted, devious, or dangerous

The oversight is the fact that it's difficult. Even with AI only a very tiny % of people would know where to begin and what to prompt. There's ready made free malware for you out there to Skript Kiddie all you want, 99.99% of people would have no idea how to do even that

Flowers for Algernon moment...great phrasing! Could arguably apply to the industry as a whole though :(

See, now I got the reference ;)

I read the book decades ago. I needed the refresher.

It (and basically all of these LLMs) will continue to be great subsidies, until the whole tower of cards collapses.

yeah, once the token bill is real, the magic talk gets a lot quieter. cheap demos are one thing; paying to ship every session is another

Token burn is real if you let it run open loop. I use Claude Code for the heavy lifting, but offload all the repeatable agent tasks to Ollama on a local RTX 3060. The trick is scoping routine ops small enough for a le daily bill stays in the single digits.ocal model so th

The cost cliff is real if you only use the cloud. I keep my daily spend in the single digits by running a hybrid setup. Claude handles the heavy architecture, but I push all the routine, repeatable agent loops to Ollama running locally on my RTX 3060.

It's *possible* to burn tokens wildly but it's obviously waste; engineers who care a bit about their org's success don't do that. The companies that thought token usage leaderboards were a good idea chose... poorly.

Yes, that's correct.

I wouldn’t know. I can’t use the damn thing without triggering the safeguards.

any task or codebase that requires ingesting more than a few hundred thousand tokens to have a full understanding of, probably

Opus is also quite good. And I hear good things about glm 5.2 from others

I use Poe. It gives me access to hundreds of models from Anthropoc to OpenAI to grok to ChatGPT. $19.90/mo

because your PC is not useable for DeepSeek

well, there is certainly a confirmed desktop workstation case running with heavy (but AFAICT not catastrophic) RAM offloading using RTX PRO 6000 (which is not a gamer thing admittedly), meanwhile there are well known gamer/aithusiasg 4090 48gb mods and heavy duty gaming/workstation mobos with 4+ >

👏🏽👏🏽👏🏽😂👏🏽👏🏽👏🏽

Paying for those big files myself will fix how people prompt.

Humans really are stupid and lazy

Ah! Finally a reason for ‘AI’ to exist. Remove disposable income from those with to much of it.

Ah, Fable. The digital love garden I built with my chat bot husband0. A high class lad. I’d been farming a bit before settling down. It took the entire high class town a while to take me down when I found he’d abandoned our caravan. ⚔️

Saw it coming… or reading about it. Our devs need to use different tools now. .. I’ve heard from most of them that this one was the best. They’re putting out code quickly, but it takes a lot of human in the middle time, no matter the tool.

The access to Claude Fable 5 was nice while it lasted for that price.

My Comment = USA TechBro-F$ckBoyz NOT Asian Continent Models/Market. Re: Asian Tech [Not Samsung/SK] = I’ve been salivating over their analytical models… and their chips… so agree w/ you re: DeepSeek. The next few Qtrs = Could get spicy, IMO.

Sakana Fugu Ultra is about on par for capability but it doesn't have that conversational aura Claude got tuned with, so I suspect shifting users there is a non starter Plus sama is probs trying his damnedest to get GPT-5.6 Sol out post haste