pangolin

eng lead @specialmech @consensys

[pangolin @0xpangolin]

there's a 2020s technology that has radically changed the lives of billions of people, but it's not AI

[pangolin @0xpangolin]

so they've 3x'd their code output since opus 4.5 only to ship increasingly worse models

[Anthropic @AnthropicAI]

Today, Anthropic engineers on average ship 8x as much code per quarter as they did compared to 2021-2025.

[pangolin @0xpangolin]

beginning to sell AI stocks. Reasons:
- I'm seeing genuine "this time it's different" euphoria, which i haven't seen since 2021 crypto
- claude 4.8 is the same or worse than 4.7, 4.6 and 4.5
- spacex IPO

[pangolin @0xpangolin]

Qualitatively I have seen zero progress on this. They are constantly, relentlessly, confidently wrong about the simplest things. I still get a lot of value out of them but I don't currently see a path to superhuman intelligence in any domain without a tight verification loop

[pangolin @0xpangolin]

the AI progress fallacy:
"ChatGPT is totally wrong 20% of the time, but that should be fixed soon"

Also see:
"Tesla Autopilot basically works, should be fully autonomous soon (2017)"
"Now we've beat Atari games, we can simply apply the same RL algorithm to the real world (2015)"

[pangolin @0xpangolin]

we spent billions of dollars making everywhere wheelchair accessible and now we're spending billions of dollars figuring out how to give robots legs

[Brett Adcock @adcock_brett]

F.03 can now walk up/down stairs purely using it's onboard camera perception

Our robots now walk from manufacturing when built to HQ

This is trained end-to-end with reinforcement learning in simulation

[pangolin @0xpangolin]

@atmoio the simpler interpretation is that LLMs are extremely bad at writing

[pangolin @0xpangolin]

mixed reviews on goodreads

[Third @thirdmetax]

Mr beast reveals he made his first 250 employees read ‘The Goal’ so they understand when he says Bottleneck

“I made my first 250 employees read it it helps get everyone on the same page of when I say bottleneck I use the word bottleneck quite a bit especially when filming”

“if I tell you you’re the bottleneck to the production to me that’s a very very serious sentence”

[pangolin @0xpangolin]

confused why people say this. AI writing IS objectively bad. It reads like a transcript of someone talking mixed with marketing copy. It's been finetuned into this strange one-size-fits-all uncanny valley prose that you literally never see in human writing anywhere

[hilary gridley @yourgirlhils]

today’s amazing new AI-designed artifacts will look like slop in a month, once everyone learns to recognize the patterns the model falls back on. like AI-generated writing, the output isn’t objectively “bad,” (in fact it is often technically quite good), but once it becomes predictable, it reveals itself as recognizably “AI.”

this is undesirable because it exposes two separate skill issues:

1. the person lacks the design (or writing) taste to realize their work reads as obviously “AI”
2. they also lack the prompting skill to steer the model away from its default patterns

this is why there will always be a signaling arbitrage opportunity in keeping a human in the loop for creative and many kinds of knowledge work, no matter how good the tools/models get

[pangolin @0xpangolin]

@minhsmind google sells TPU capacity to anthropic. They make more profit from those subscriptions than anthropic does

[pangolin @0xpangolin]

idk man... every 6 months they say the new model is dangerously good.

anthropic have been using this one for months, claude code is buggy slopware and they have 148 open engineering roles

[pangolin @0xpangolin]

@toly @alessandrod you are describing jito

[pangolin @0xpangolin]

demis is the only AI CEO not larpmaxxxing. Deepmind are the only major AI lab that has made true scientific progress with AI. He has a nobel prize and is an actual genius. They are the only AI lab doing more than just scaling LLMs. It will be obvious in hindsight when they win

[Steve Hou @stevehou]

“Hassabis administered a subtle test on him. The two men discussed the potential of AI, and Zuckerberg expressed appropriate excitement. But then, as the dinner continued, Hassabis brought up other hot technologies: virtual reality, augmented reality, 3-D printing. Zuckerberg sounded equally excited about all of them. ‘That told me what I needed to know,’ Hassabis said.”

[pangolin @0xpangolin]

@johnrobertreed the term always caused confusion because the topic is complex and <1% of people who talk about it actually understand it

[pangolin @0xpangolin]

@alranpe almost all blocks are sequenced offchain by jito anyway so the protocol design really doesn't matter for tx sequencing or censorship

[pangolin @0xpangolin]

@mert they shipped a blog post

[pangolin @0xpangolin]

multiple concurrent proposers doing whatever jito tells them to do. Truly game changing stuff

[Anza @anza_xyz]

Introducing Constellation: a multiple concurrent proposers protocol for Solana. With proposers operating on a 50ms cycle, Constellation gives Solana the fastest protocol enforced economic tick rate of any blockchain.🧵

[pangolin @0xpangolin]

@MaxResnick impressively wrong on almost every detail here

[pangolin @0xpangolin]

the true societal benefit of LLMs is their ability to oneshot the specific type of narcissistic midwit who traditionally fails upwards into positions of power

[pangolin @0xpangolin]

they don't want you to know that it's trivial to achieve excellent cardio health. You can get 50+ VO2max doing a 15 minute run once a week. Cardio fitness is gatekept by weird running addicts

[pangolin @0xpangolin]

these people were annoying as hell in real life jobs

[taoki @justalexoki]

i think its a little sad to watch all these people who took pride in the art of programming slowly die away as they realize bit by bit that its actually over

[pangolin @0xpangolin]

FOCIL is a much better solution than MCP. If transactions are not fully private they will always be sequenced by an auction winner. MCP just makes the auction more complex but ultimately changes nothing. Things like FOCIL are good in the short term for censorship resistance and long term we need fully private txs