Peter Cooper
@cooperx86
Followers
25K
Following
22K
Statuses
81K
Father of three and GitHub user #118. Programming, LLMs, 3D printing, and piano/music theory. Currently: doing LLM inference experiments.
United Kingdom
Joined December 2006
I've turned my MLX LLM inference thing for macOS into a pip package: # pip install infermlx import infermlx.infer as infermlx model = infermlx.Model.load_model() model.generate("Tell me a joke") .. and you're off to the races. All you need is Python and a Mac.
1
1
6
GitHub has added Nerd Fonts, box drawing glyphs, and all sorts of goodies in the new release of their Monaspace of monospaced fonts:
0
0
1
Saying LLMs "just predict words" is not inaccurate, but as needlessly reductive as saying Conway's Game of Life "just flips bits" while ignoring the patterns and behaviors that emerge.
0
0
2
@renegadesilicon @tmdanis Does this suggest that if models were able to continually fine tune themselves in situ, we could get even better results?
1
0
0
Further, so-called "aha!" moments are the brain's reinforcing emotional response to detecting, integrating, and synthesizing a novel pattern from prior knowledge. LLMs lack the emotional response, but they can still do the latter part.
Can humans reason? In this paper we show evidence many humans simply apply heuristics they've been exposed to over the course of their lives without deeper consideration. In conclusion, humans don't seem to reason and only copy reasoning patterns from their training data.
0
0
1
@dontdelay @AlanJLSmith Sure but that could be changed in legislation surely without much public outcry. Like with “hobby” income, taxable but losses only deductible up to the amount of any winnings?
0
0
0
@jhooks This is how I feel when industry luminaries start spouting crap like "say no more often", "stop having meetings" and similar stuff. Great, if you have cred and are successful. Not so great as a newcomer to the workforce 😂
1
0
5
@jaymcgavren This just popped up in my feed(!!)
In the late 1960s, CIA disguise chief Tony Mendez developed an unusual escape tool called a "scrotum concealment device." This latex scrotum was designed to discreetly hide an escape radio in a spot unlikely to be detected, even during strip searches. Sadly, the device was never advanced beyond the prototype stage.
0
0
0
@jaymcgavren It boggles my mind that anyone ever thought a single cat would make a good spy out in public. They're way too unpredictable!
0
0
0
@liran_tal @elevenlabsio Might throw some JavaScript documentation into it and see what nonsense it comes up with 😂
1
0
0
@DavidEtherton1 I think we'd first need to define "original thought". Is a 5 year old inventing an animal with ten legs, laser eyes, and a funny name having an original thought? I mean, it's just stuff they've seen before merely synthesized in a unique way..
0
0
0
@BongoEstente @lewis_goodall True, once the UK declared independence from the EU it finally got its own government, military, and laws and no longer had to be in that pesky Schengen zone 😉 <whispering at the peanut gallery replying to you> the UK always had sovereignty..
0
0
1
@jmitch @jasonleowsg Yeah that’s bizarre. Maybe they could up it with the much cheaper o3-mini. I use Aider instead and $200 would get me probably ten thousand uses a month 😀
1
0
2
@burkov Strong disagree, but concede it depends entirely on how you define the terms of "generate" and "information".
2
0
12
@awnihannun I've been doing some similar stuff today but in a Llama-only heavily boiled down mlx-lm variant (solely to reduce complexity for my less experienced brain!):
0
0
1
You can't use dynamic temperature during the non-thinking stage as the HTML goes totally off the rails. But during thinking, the temperature fluctations make it more creative in its planning!
0
0
0
You can also STOP it thinking after a defined amount of tokens which is a good way to stop thinking loops (common in the low parameter versions of DeepSeek).
0
0
0