John Freier

Entrepreneur, GenAI Addict/Enthusiast, Web Dev, Statistician, Dirt Biker, Willed Biohacker

[John Freier @johnf_ucb]

@logicus Helps immensely when you have guidance / teacher / mentor! But yeah, I was laughing the other day about how much of a pain it was with R (stats lang) - no IDE, no git, horribly unhelpful error messages. BUT, searching for bad error messages in conjunction with “R” …PITA

[John Freier @johnf_ucb]

@mattshumer_ Get started

[John Freier @johnf_ucb]

@mattshumer_ First serve!

[John Freier @johnf_ucb]

@DualAcies @Hesamation The window of high system loads reminds me of the early internet days with Shared Hosting providers like @PowWeb and the issue of “noisy neighbors”

[John Freier @johnf_ucb]

I actually have aggressive rules added in for further emphasis against anything remotely sycophantic, and that if anything *actually* and legitimately is unique or impressive or whatever, a comment may be made if and only if it is strongly backed up with rock hard evidence. Works great.

[John Freier @johnf_ucb]

This happened to me as well and at a very notable point. With an interesting (well, interesting if it’s the first time to experience this) end result. TLDR; CC makes up for it (see last line of this message)

1. I think I was nearing 300k tokens / 1m (I’ve definitely felt the degradation of quality when at greater token counts than 200k - no hard evidence other than this particular even but..when you’re one to be cognizant of those things and you use the tool day in and day out, most people should be able to sus this out..)

Higher percentage context window usage leading to reduced performance is also a known pattern..

2. Claude ended up taking me down this bs implementation route which ‘had the looks-good vibes’ despite being completely useless data processing and value claims. Honestly, I flipped shit at it ha. Then, despite want to just thrash things, I dropped a bunch of the most recent messages into a fresh CC session and just said something to the tune of ‘what in the actual f**K Claude!? Go sequential think to 100.’

I was not really expecting anything back, just… feeling pathetic by the fact that I was mad at an AI, especially since it would then be a totally independent and separate instance; so I felt I’d tell it off by telling it to sequential think to 100 (with really no explanation other than the pasted messages and my one line of anger).

It was the closest I’d get to telling it to go endlessly spin its wheels and think about what it has done!! (yeah… screw you computer, run an endless while loop that writes junk all through your filesystem! ..damn..that’s my filesystem..).

Then I leave to address my clearly bottoming-out blood sugar levels.

!! Turns out, the new session analyzes the pasted context really well and explains how Claude[session 1] was, in fact, doing all the deceptive things I was pissed about.
And then I told it how screwed I was now on my project’s timelines, and Claude kicked into gear and just started tearing through things based on the original pursuit and crushed everything in ~45min.. (as in, what had been a multi day WIP..)
#claudecode #context-window-usage

[John Freier @johnf_ucb]

@transcended @cryptohuntz @Geiger_Capital #HOTL

[John Freier @johnf_ucb]

@tymrtn @1RustyMac @grok @grok ELI5 the white paper

[John Freier @johnf_ucb]

I appreciate the transparency. Though, keep in mind that (or if this wasn’t known, here you go ;-) ), LLMs have various variables which are set during the output and they affect how the LLM operates with respect to probabilities of various outputs. As such, if you were to do this many times, you will inevitably get a different result actually.

This is actually one of the ways in which the security folk “red team” LLMs into divulging information they shouldn’t: literally by repeating a question / prompt over and over and voila.

Just thought I’d share, still appreciate your dedication to presenting your approach with maximum transparency!

[John Freier @johnf_ucb]

Thank you, I was wondering how many replies I’d have to read before someone informed the thread of this… the conversation thus far reads as if LLMs are deterministic.

Though I don’t think Elon is necessarily acknowledging he/they are steering in a particular direction.

Steering is going to be to LLMs as bias is to humans.

[John Freier @johnf_ucb]

@thecsguy @kimmonismus Seriously..

[John Freier @johnf_ucb]

@manpreets7 @MatthewBerman If you really want it just slower, I suppose you could add in some lsp servers and whatever random and sluggish hooks suit your fancy..

[John Freier @johnf_ucb]

@johnrushx @nixeton Well, on the bright side failing fast is a good thing ;-)

[John Freier @johnf_ucb]

@harrisonwang13 @kimmonismus Claude is that you??

[John Freier @johnf_ucb]

@kimmonismus So you’re saying it knows how many r’s are in strawberry?

[John Freier @johnf_ucb]

• What if confusion is a kind of map?

• ⁠** What if trauma isn't pain but corrupted memory code? **

• ⁠Why do some things break when you try to define them?

• ⁠What's the shape of a belief just before it collapses?

• ⁠** What if every emotion is a different type of logic protocol? **

• ⁠Can a system evolve if its only rule is: respond to contradiction?

• ⁠What's on the other side of a paradox if you never stop walking into it?

[John Freier @johnf_ucb]

@flutterwhat @skirano Multi agent debate and seek consensus

[John Freier @johnf_ucb]

One day to crush through my $1k Claude Code for Web credits woo #claudecode

[John Freier @johnf_ucb]

@elonmusk The lolz of not getting the invite to the EV event by Biden Administration just continue to pay dividends

[John Freier @johnf_ucb]

@_junaidkhalid1 @cerebras Raises hand