> [Humans] can be really bad about just trying things and seeing what works. If you’re not watching all the tool calls you may not see it, but it’s kind of scary to watch them just bump into wrong decisions and backtrack.
3,062 karma · joined December 1, 2012
To give more anecdata — Union Station in Toronto has trains that leave in all directions. The ones going East and West tend to form queues on the platform at rush hour. The train heading north-west to another city tends to form mobs that rush the door upon opening.
Lots of reasons why, but geography and SES may not factor as much. Could be something as simple as supply/demand of free seats.
If frontier AI is load-bearing, be thankful you're able to spin up comparable models (face it — 99% of us aren't writing novel software) for a fraction of the cost it would be for a third-party to come in and mop up the mess.
My model weights don't change unless I change them.
Yeah, expecting the world when all you have is a 8GB graphics card? You're going to be disappointed.
16GB is table stakes (IQ3_XSS). 32 GB is better.
No need to mess around further.
One can (and probably should) make the argument that if your initial/incremental belief was incorrect, one should start again with a fresh prompt or roll back the tree and prompt again, disposing of the now-incorrectly generated code.
I catch myself doing this all the time. My model (local) isn't nearly as fast as something cloud-based, so there is a real incurred cost of time that is hard to shake. Very often I'll need the model to reconsider and rewrite the plan after it's already done quite a bit of work. Rolling back the code is tough, but I should resort to it more often.
I'm sorry to be the bearer of bad news: human coding has not improved a lick since then.
Take the most recent qwen and deepseek models with offloadable n-grams, which function (both in name and vaguely in capability) like human memory "engrams".
In my opinion, 98% of the work most devs would send to an AI can be capably achieved with a local model and a frontier-level model is overkill.
The goalpost moving feeds right into Anthropic and OpenAI's interests.
Though keep in mind not being beholden to shenanigans from said cloud companies (and interference from government entities!) is definitely worth something intangible.
I can cherry pick stats too.
The other day I heard mention of someone paying $200/mo for Claude Code.
At those rates my local LM setup pays for itself in a single year.
It seems to be a low-effort compromise to running a VM or separate machine just for the agent, and I am still able to audit the model's work in my IDE.
Uninstalling the app fixed it up.
> It was intended that when Newspeak had been adopted once and for all and Oldspeak forgotten, a heretical thought — that is, a thought diverging from the principles of Ingsoc — should be literally unthinkable, at least so far as thought is dependent on words.
Getting to the point where I was able to run a 30B model required $500 in memory.
This does happen without AI already. My local shawarma joint has a deal of the day poster of a chicken shawarma that's obviously not theirs. Next to a bottle of coke.
The poster is sponsored by Coke and the restaurant just slapped on a price.
The same arguments pop up when new therapies are unveiled for various sightedness disorders, hearing impairment, etc.
How _dare_ you impugn upon my identity that has been reinforced through my upbringing constantly? Etc.
Not everybody measures weight in pounds, and to the primary readership of the BBC, stones and kilograms mean much more than weighing pounds.
Source: Canadian who understands neither kg or stones.
Artificial inflation to recoup R&D.