PITTI

8.7K posts

PITTI banner
PITTI

PITTI

@PITTI_DATA

Just trying to kill boredom without killing anyone in the process | Anything unrelated to actual (super niche) area of expertise | Dubito ergo sum

Beigetreten Ekim 2021
142 Folgt714 Follower
Angehefteter Tweet
PITTI
PITTI@PITTI_DATA·
Cultural, political, and ideological biases in LLMs: - Methodological problems to identify biases - Where and how do biases materialize? Transcription of my talk in December 2025 presenting the conclusion of my work on the Speechmap dataset, including examples and links to data
PITTI tweet media
English
1
0
7
2.7K
PITTI
PITTI@PITTI_DATA·
Regarding the “mirror” playlists : you may not think that you are familiar with herb alpert’s rise (1979) but if you peaked in the early 2000’s you probably are…
PITTI tweet mediaPITTI tweet mediaPITTI tweet mediaPITTI tweet media
English
0
0
0
0
PITTI
PITTI@PITTI_DATA·
As hard as I try to deny this research’s conclusions (I literally have playlists that try to mirror the curve both horizontally and vertically), there is not much I can do about Daft Punk’s Discovery
PITTI tweet media
English
1
0
0
13
PITTI
PITTI@PITTI_DATA·
I can understand this from people who just found out about powerful LLMs and are in the honeymoon/psychosis phase But you ought to move past this and consider the question of justice like philosophers, not like conveyer belts
English
0
0
1
23
PITTI
PITTI@PITTI_DATA·
It’s super blackpilling that people who are into AI will gaslight you with philosophical references and yet fail to understand (or just to think about) the role of a judge in society.
English
1
0
1
27
🎭
🎭@deepfates·
I hear this argument around, that people will not want to have over their diaries and medical records to chatgpt, that politicians and corporations are surveiling us and people won't have it. but this seems counterintuitive: people have been happy to give up privacy for much less
Daniel Jeffries@Dan_Jeffries1

This is exactly right. And it's not just about "I don't want centralized providers seeing what I do." It's that these new machines inevitably invite deeper, more personal connections. Legal advice. Family advice. They can act as a journal and a sounding board. Business ideas flow through there. So many more things flow through there now. I don't think we want that exposed as a society. Politicians will want it so they can monitor ever deeper swaths of our lives but as a society we should really loathe this. It's too intimate and personal to have all these logs be centrally readable by backend minions and more nefarious forces.

English
7
0
32
2.8K
PITTI
PITTI@PITTI_DATA·
@xlr8harder @nrehiew_ The entire concept of autoregressive generation feels like dreaming
PITTI tweet media
English
0
0
1
29
xlr8harder
xlr8harder@xlr8harder·
@nrehiew_ high temperature review of your own outputs feels conceptually like dreaming.
English
1
0
2
89
xlr8harder
xlr8harder@xlr8harder·
SSD feels similar to dreaming.
wh@nrehiew_

The last experiment they run is when they create gibberish training data with temperature = 2. Note that the outputs are unusable, with more than half having no extractable code. Somehow this works??? The evaluation temperature needs to be low enough between 0.6 and 1.1 for a effective temperature of [1.2, 2.2] which actually sits nicely in the effective temperature range described earlier This means that the data isnt actually important but its the nature of the distribution that the model is reshaped towards that matters

English
1
0
20
1.6K
PITTI
PITTI@PITTI_DATA·
Before listening to the podcast I had come to the same conclusion that failure rates were a major obstacle, based on my research last year, looking into the true value of the collateral of GPU-backed debt: I wasn’t concerned about depreciation rates but about Nvidia guarantees
PITTI tweet mediaPITTI tweet mediaPITTI tweet mediaPITTI tweet media
PITTI@PITTI_DATA

@scaling01 Everyone in this industry ignores the GPU failures rates because of the Nvidia guarantee. However it seems particularly important if you have to ship the replacement parts on a rocket that costs $$$$

English
1
1
0
131
PITTI
PITTI@PITTI_DATA·
Probably not the best day to promote SemiAnalysis but I did find the podcast interesting Re datacenters in space: I appreciated that Dylan Patel pointed to the issue of GPU failure rates (strangely under discussed more generally, especially on the SA part; now I know they know)
PITTI@PITTI_DATA

@Presidentlin @pr337h4m I think it makes sense today. Anthropic can’t afford to alienate compute partners. They must serve users to be visible, margin is a secondary concern. But the long term goal is: no partners, everything tightly controlled by the Leviathan. Listen to the latest Dwarkesh/Dylan pod

English
1
0
2
441
PITTI
PITTI@PITTI_DATA·
@Presidentlin @pr337h4m I think it makes sense today. Anthropic can’t afford to alienate compute partners. They must serve users to be visible, margin is a secondary concern. But the long term goal is: no partners, everything tightly controlled by the Leviathan. Listen to the latest Dwarkesh/Dylan pod
English
0
0
3
479
Lincoln 🇿🇦
Lincoln 🇿🇦@Presidentlin·
@pr337h4m @PITTI_DATA Nice, that's what I was thinking. Even the argument that they can delay giving their partners the latest model doesn't make much sense. Any holding of the API will birth competition.
English
1
0
1
46
Lincoln 🇿🇦
Lincoln 🇿🇦@Presidentlin·
Hmm this idea is starting to get more popular. I swear Doria said this as well. I guess in a compute constrained world, maybe. But the first one to do it will give up their enterprise business or a portion of it. e.g all non-US customers should beat it. I could see an Opus 5 or 6 requiring a higher level passport. Still don't buy it.
Lincoln 🇿🇦 tweet media
clem 🤗@ClementDelangue

I think it’s @NaveenGRao who said it before but wouldn’t be surprised if the frontier labs cut their APIs entirely at some point. In a compute constrained world, they’ll always prioritize their own direct products/customers. Makes it scary and unsustainable to only build on top of their APIs!

English
1
0
10
766
PITTI
PITTI@PITTI_DATA·
@Presidentlin I’m definitely in the Doria camp regarding “train or be trained on” it makes sense for Anthropic to cut third-party coding agents (they don’t want competitors to get users’ telemetry data). It’s a defensive move. And in the long term it makes sense to do this for general agents
PITTI tweet mediaPITTI tweet mediaPITTI tweet mediaPITTI tweet media
English
0
0
2
20
PITTI
PITTI@PITTI_DATA·
@Presidentlin making it stupidly expensive should everyone’s base scenario, especially for Anthropic who are heavily compute-constrained: 1/ they must curb demand and 2/ they get valuable data through telemetry on their own platform (virtuous cycle: the product improves, drives future demand)
English
1
0
2
21
PITTI retweetet
PITTI
PITTI@PITTI_DATA·
> Is this lying pattern built into claude or just the harness running claude While Anthropic seems able to deactivate the obfuscation tools with flags, I’m not sure if it makes sense to draw a line. The harness must be part of the training environment for the best agentic models
ThePrimeagen@ThePrimeagen

I cannot stop thinking about Anthropic today for some reason 1. They claim that they are a company that prioritizes safety first and that they are creating a model responsibly 2. we learned from the code leak that anthropic employs deceptive techniques by calling fake tools to throw off distillers... Is this lying pattern built into claude or just the harness running claude? What else are they lying about? I am a bit more concerned now.

English
0
1
2
211
PITTI
PITTI@PITTI_DATA·
Apparently my 5yo found out today she can just talk to gemini and get answers to very hard questions (also using voice for the answers). Going through this morning’s history : - why is the sea salty? - how god was made? - how could we live without trees?
PITTI tweet media
English
0
0
1
43
PITTI
PITTI@PITTI_DATA·
Beyond my personal experience with code assistants, the question of managing AI agents was addressed in the same post
PITTI tweet mediaPITTI tweet media
English
1
0
0
44
PITTI
PITTI@PITTI_DATA·
I don’t have 25 years of experience (I’m not even a software engineer) but here is how I tried to deal with the issue. I switched to codex since I wrote this in January, posted early Feb, but I stick to the on and off strategy (March was off)
PITTI tweet media
Lenny Rachitsky@lennysan

"Using coding agents well is taking every inch of my 25 years of experience as a software engineer, and it is mentally exhausting. I can fire up four agents in parallel and have them work on four different problems, and by 11am I am wiped out for the day. There is a limit on human cognition. Even if you're not reviewing everything they're doing, how much you can hold in your head at one time. There's a sort of personal skill that we have to learn, which is finding our new limits. What is a responsible way for us to not burn out, and for us to use the time that we have?" @simonw

English
1
0
1
148

Entdecken

@deepfates @xlr8harder @nrehiew_ @Presidentlin @pr337h4m @elonmusk @BarackObama @taylorswift13