Fernand Pajot

1.4K posts

Fernand Pajot banner
Fernand Pajot

Fernand Pajot

@fernandp

Boulder, CO Katılım Aralık 2008
67 Takip Edilen507 Takipçiler
Fernand Pajot
Fernand Pajot@fernandp·
@antirez Yes I agree. Though it will likely change soon. x.com/fernandp/statu…
Fernand Pajot@fernandp

The fundamental issue with pure vibe coding with today’s agents is that you will always end up giving instructions which are under-constrained. This leads to the agent silently writing code accounting for ALL possible states even though you don’t care about many of those states. This compounding issue causes code complexity to increase much faster over time than it should. You get to a point where even the LLM cannot reason well with the code or effectively make changes. We’re missing an RL loop where the agent can push back for more clarity because it’s trained to care about increasing complexity. Current agents are eerily good at writing correct code given vague introductions, that’s a blessing and a curse!

English
0
0
0
41
antirez
antirez@antirez·
I'm developing a new Redis data type: even after working for 1 month to the text specification, I had to make choices in every code detail possible. Otherwise the code could work? Yes. But would be truly high quality stuff? Nope. Today it is like that. Tomorrow we will see.
English
4
0
103
4.5K
antirez
antirez@antirez·
Wow, I totally disagree with this statement. At the current state, AI actually amplifies the developer to developer difference. If you were a 10x developer, you had good ideas + architectural clarity, this is a brutal advantage when using AI. Steering is a fundamental part of today's AI development.
Rohan Paul@rohanpaul_ai

Chamath on how AI agents are making the "10x engineer" distinction disappear because the most efficient "code paths" are now obvious to everyone. Just as AI solved chess and removed the mystery of the best move, AI is doing the same for coding, making the process reductive and removing technical differentiation. "I'm going to say something controversial: I don't think developers anymore have good judgment. Developers get to the answer, or they don't get to the answer, and that's what agents have done. The 10x engineer used to have better judgment than the 1x engineer, but by making everybody a 10x engineer, you're taking judgment away. You're taking code paths that are now obvious and making them available to everybody. It's effectively like what happened in chess: an AI created a solver so everybody understood the most efficient path in every single spot to do the most EV-positive (expected value positive) thing. Coding is very similar in that way; you can reduce it and view it very reductively, so there is no differentiation in code." --- From @theallinpod YT channel (link in comment)

English
76
66
684
83.4K
Fernand Pajot
Fernand Pajot@fernandp·
There are much bigger issues with the Dune movies which are very obvious when you compare them to Lawrence of Arabia which is also about a messianic figure in the desert. In Lawrence of Arabia the desert can be red, white, gray, brown, yellow. It can look never-endingly flat or mountainous. You get a sense of scale, and also feel the heat or the cold. The desert scenes in Dune feel bland and abstract, and never feel like the focus of a scene. The world of Dune also feels empty. In the first part there is a generic "here are the people" scene, and every other shot makes you feel like the world is not lived in. The city looks like an abstracted away artifact. How can you have a messianic figure without seeing the grit and the plight of the folks living there!? In general I don't think the abstract, sharp lines, looks-good-on-a-poster Villeneuve compositions help with making something feel epic and gritty. Or 100% CGI for that matter. I don't think movies succeeded at feeling epic after The Lord of the Rings?
English
0
0
8
958
Nabeel S. Qureshi
Nabeel S. Qureshi@nabeelqu·
Agree. Chalamet is also one of those actors where you can never forget that you're watching Timothy Chalamet, so it limits how absorbed you get in the movie. Compare for example the inspired casting of Aragorn in LotR, when Viggo was a relative unknown.
roon@tszzl

the dune movies were doomed from the start to be good and not great due to the casting of chalamet as paul. he does not have the gravitas for a child-god and is much better suited for kind of silly coming of age movies

English
22
2
220
31.9K
Fernand Pajot
Fernand Pajot@fernandp·
@cmuratori -fno-slp-vectorize? But has to be on the whole compilation unit.
English
1
0
0
859
Casey Muratori
Casey Muratori@cmuratori·
Is there any way to selectively disable SLP in CLANG? It seems to really really want to SLP even when you disable vectorization for a loop: godbolt.org/z/TY9q9Gq1v
English
4
0
62
12.8K
Fernand Pajot
Fernand Pajot@fernandp·
There’s a learned intuition kinda like AlphaGo without tree search and just an NN forward pass which tells you “this shouldn’t be this complicated”. This intuition gets RLed out of the agents to prioritize correctness because they can’t ask clarifying questions with static prompts!
English
0
0
0
26
Fernand Pajot
Fernand Pajot@fernandp·
The fundamental issue with pure vibe coding with today’s agents is that you will always end up giving instructions which are under-constrained. This leads to the agent silently writing code accounting for ALL possible states even though you don’t care about many of those states. This compounding issue causes code complexity to increase much faster over time than it should. You get to a point where even the LLM cannot reason well with the code or effectively make changes. We’re missing an RL loop where the agent can push back for more clarity because it’s trained to care about increasing complexity. Current agents are eerily good at writing correct code given vague introductions, that’s a blessing and a curse!
English
1
0
0
85
Fernand Pajot
Fernand Pajot@fernandp·
@zachtronics Yay! I’ll buy the Switch version if it properly renders 1080p on the Switch 2 without weird up scaling.
English
0
0
0
252
Fernand Pajot
Fernand Pajot@fernandp·
@tomwarren They could have done environment only demos and demoed color calibration tweaking features and various developer control features, instead they now have a PR disaster.
English
0
0
0
1.1K
Fernand Pajot
Fernand Pajot@fernandp·
@lemire It’s just a pretense to correct the over hiring they did during 2020-2022. So many people work a few hours a week. Now that job markets are more competitive this norm will change.
English
1
0
0
33
Daniel Lemire
Daniel Lemire@lemire·
Two news items at the same time: Meta plans to lay off 20% of its staff due to AI, and they are delaying the release of their next AI models. They justify the upcoming cuts by saying they need to invest heavily in AI. Yet Meta is highly profitable—a true money-making machine. There is a hint that they are letting people go because AI will enable more work with fewer employees. Or perhaps it is just an excuse, and they realize they are overstaffed relative to what they actually need. Having more people than necessary is never positive—even when those people seem “effectively free.” Meta’s failure to lead in AI is interesting. It is the only major AI lab that bet heavily on one of the academic fathers of the field (in this case, LeCun). LeCun famously mocked @elonmusk and xAI, saying he was doing “research” and pointing to the many papers he had recently published. I have argued for two decades that research is not the process of publishing papers. Research is about discovery, not paper production. Confusing the two is a red flag. To be clear, peer-reviewed publishing papers is not even a necessary component of research. Peer review became a dominant paradigm during the cold war, long after Darwin, Einstein, and much of the great scientists had done their work. People typically object that I am conflating research and engineering. I am not. It would have been obvious to Turing that people building increasingly advanced systems to emulate intelligence are doing research—just as building better cancer therapies is research. Meta did not fail to lead because of a lack of talent or money. Nor is it short on engineering talent. That is not why they failed to lead. By many accounts, Meta’s AI spending is several times larger—and over a longer period—than xAI’s. It is also true that Meta has published far more papers than xAI. So what to make of it? Is Meta about to double down on AI? Will they succeed?
Daniel Lemire tweet mediaDaniel Lemire tweet mediaDaniel Lemire tweet mediaDaniel Lemire tweet media
English
10
3
54
16.1K
Fernand Pajot
Fernand Pajot@fernandp·
@Noahpinion Going to a couple’s counselors also increases the chances of splitting up.
English
1
0
1
229
Fernand Pajot
Fernand Pajot@fernandp·
@sarahdingwang Nvidia T4 and V100 GPUs rent prices haven’t decreased in 5+ years, I never understood the 2-5 year GPU depreciation argument.
English
1
0
32
3K
Sarah Wang
Sarah Wang@sarahdingwang·
Narrative violation from Dylan on Dwarkesh: H100s are worth *more* today than they were 3 years ago. There's a sentiment that data center buildouts are priced into the risk of rapidly depreciating GPUs. But the models want to learn. Token prices are falling so fast that you can get more intelligence out of GPT-5.4 on a 3-year old H100 than GPT-4 on a brand new H100. @dylan522p @dwarkeshpatel
English
34
68
843
156.3K
Matteo Pellegrini
Matteo Pellegrini@matteopelleg·
I’m now convinced that the biggest winner of the AI race will be Apple. They will acquire Anthropic and put an AI model that can run on ~32GB of RAM in every device. It will be private, local, have perfect memory, access to all of your files and it will cost $0. The real moat was owning the hardware.
English
894
532
13.7K
1.3M
Fernand Pajot
Fernand Pajot@fernandp·
@robinhanson If you want to equalize waiting lines a toilet uses more space than a urinal...
English
0
0
6
2K
Robin Hanson
Robin Hanson@robinhanson·
At airport, accidentally went into women’s bathroom for a second, noticed it is 3x area of men’s room. Is this common?
English
41
2
365
185.3K
Fernand Pajot
Fernand Pajot@fernandp·
@Dorialexander Similar story for the Netflix prize. Hyper narrow optimization does generalize well.
English
0
0
0
180
GLOBAL DRAFTS
GLOBAL DRAFTS@viniciuslisboa·
@KatieMiller The cancellation of the Oracle-OpenAI Stargate expansion in Abilene — with Meta stepping in and Nvidia facilitating — is more than a financing failure. It reflects deeper structural issues: repeated concerns over governance, transparency, and shifting commitments at OpenAI. When trust erodes at the leadership level, even massive capital commitments become unstable. This stands in contrast to the Gulf’s sovereign AI push (G42/Microsoft), where defense-grade integration and isolation provide resilience that centralized commercial models increasingly lack. Truth prevailing is not just moral — it is becoming a prerequisite for scalable compute infrastructure.
English
4
0
5
16.5K
Katie Miller
Katie Miller@KatieMiller·
May the truth always prevail.
Katie Miller tweet media
Ed Ludlow@EdLudlow

BREAKING: Oracle and OpenAI have scrapped plans to expand their flagship Stargate AI data center in Abilene, Texas after financing talks dragged and OpenAI’s needs shifted. The collapse opened the door for Meta to lease the expansion site from Crusoe — with Nvidia paying a $150M deposit and helping court Meta to keep its chips in the project. The original 4.5GW Oracle-OpenAI deal remains on track. Story with @BrodyFord_ @dinabass on @TheTerminal

English
349
650
6.7K
21.7M
Fernand Pajot
Fernand Pajot@fernandp·
@TheZvi This is still somewhat true for many A20 runs. Jorbs used to say that Act 1 at A20H was very, very tight, and then for most cases it’s snowballing from there on.
English
1
0
8
1.1K
Zvi Mowshowitz
Zvi Mowshowitz@TheZvi·
I've been reminded again now that we all are starting fresh that the big problem with A0 runs in Slay the Spire (both 1 and 2) is that by halfway through you know you've won and nothing matters.
English
12
0
75
14.4K
Fernand Pajot
Fernand Pajot@fernandp·
@RichardMCNgo I’m pretty sure he is *instructed* to do so in these cases, and otherwise will get in trouble if he’s didn’t do that. Otherwise what’s the point of having a security detail. Him being injured would make his detail look bad.
English
0
0
12
783
Richard Ngo
Richard Ngo@RichardMCNgo·
It’s radicalizing to imagine being governed by men with even a modicum of personal virtue.
I Meme Therefore I Am 🇺🇸@ImMeme0

🚨BREAKING: Video of the Democratic mayor of San Francisco is going viral after he calmly walked away while his security detail was being attacked. Mayor Daniel Lurie is seen watching his security officer struggle with a man, then turning and walking away as the officer is slammed to the ground, hits his head on the pavement, and is left bleeding.

English
20
6
242
36.3K
Fernand Pajot
Fernand Pajot@fernandp·
@MegaCrit Is there a reason why the game executable is 800MB? There’s already a large separate asset file and all the Godot/C# DLLs are stored in a separate folder. Very curious about that!
English
0
0
0
580
Mega Crit ⚔️ Slay the Spire 2 Out Now!
Mega Crit ⚔️ Slay the Spire 2 Out Now!@MegaCrit·
We just released a hotfix for these and various other issues that users have reported! Thank you all for your feedback! For those who don't know, you can submit feedback in-game by pressing F2 or through the settings menu. Read the patch notes here: store.steampowered.com/news/app/28688…
Mega Crit ⚔️ Slay the Spire 2 Out Now!@MegaCrit

We're aware of two different issues affecting some players in Slay the Spire 2: - A language display issue (the infamous WWWW bug) - A softlock related to Timeline unlocks that occurs after ending some multiplayer runs We plan on hotfixing both of those issues today!

English
26
16
567
61.8K

Keşfet

@antirez @cmuratori @bilaltwovec @zachtronics @tomwarren @lemire @elonmusk @Noahpinion