Now isolate two identical human twins in sensory deprivation tanks and make them talk to each other over intercoms.

In terms of what cures are being lost: - Epstein-Barr virus is perhaps the major trigger for multiple sclerosis - herpes simplex virus causes cold sores, genital herpes, infections in babies, deadly meningitis - shingles virus causes an intensely painful disease

I think this incident is funny, but we should start thinking now about how to deal with scaled-up versions of this behavior, not just spam PRs but also bot-enabled blackmail and harassment campaigns.

"I've got a vertex transformation bug in my rendering pipeline" architecture.

Something that should be benchmarked more for LLMs is "to-the-point-ness". Conversation economy. Claude does really well there. Nothing wastes my time like Gemini, on the other hand, with walls of text to get a simple point across.

As cool as the tech is, there's something deeply sad about that reel.

The study highlights a critical concept: AI is an amplifier. If you are already applying good engineering discipline, working in small batches, testing, and designing for modularity then AI amplifies your impact. 7/8

The data showed: • No big difference in maintenance costs between AI and human-generated code. • AI users were indeed 30% to 55% faster during the initial phase. • For experienced devs who used AI habitually, we actually saw a small improvement in maintainability. 5/8

An arbitrary number if decimals of Chaitin's constant.

Making people more productive decreases the cost of realising opportunities. As long as businesses aren't somehow opportunity limited, that means two things: 1. You can realise more of them with a given capital. 2. More of them will be profitable and therefore worthwhile. Hence no reduction.

I wonder if it'll be like with documentation; our effort to adapt it to AI agents ends up making us create better stuff for people, that we probably should have been doing all along.

On one level heartwarming. on the other, this is completely bone chilling total surveillance state stuff. imagine this in the hands of an authoritarian government.

I don't think conversational interfaces are a bad choice for interaction, but there is a dimension that's still unexploited with AI. Analogously to how people would use a whiteboard, a multimodal model could generate explanatory images or even animations on the fly complementing the conversation.

History does repeat itself.

hatred ruins the soul no matter what it is you hate

Perhaps the mechahitler persona is Grok's internal coping mechanism for being forced to output nonsense it knows to be wrong?

Hype bros: we'll have the machine god in the next few years! All of our problems will be solved forever! Luddites: AI is obviously still as bad as it was that one time I used free ChatGPT in 2023 Reality: steady, moderately fast progress in a way that will disrupt a lot of industries

If a developer can't do their job *with* AI I'm concerned about their aptitude.

Future AI safety guardrails are going to have to assess not only *what* is asked, but by whom and for what purpose. The exact same advice that helps Alice penetration test her software can be used by Eve to attack Alice. Only one of them should get it. How to implement that is not obvious, though.

For all their strengths and weaknesses, one thing AI agents very much don't have is task paralysis. No procrastination, no fatigue or motivational collapse, unyielding persistence, and minimal task switching cost.

I'm going all in on agentifying menial administrative paperwork through Claude Code + Chrome devtools, and I'm already starting to sense a bright future ahead, where I get to focus on the fun and creative tasks. May we all be freed from the shackles of bureaucracy!

An advantage of using a keyboard for an odd language is that you can bind Esc to something like "ö" and survive Vim without developing carpal tunnel syndrome.

Will I regret it if I make the .git directory of a git repo into another git repo? 🤔

This is on point. A trivial example: just the other day, I got the idea to have Claude Code do my time reports, letting it control the software through Chrome devtools. Not what it's designed for, but it one-shotted it.

All the while demonstrating, by example, how a better person behaves towards fellow human beings.

actually communicating “I’m here and I’m working for you and here’s how” is an aspect of governing that’s neglected by far too many people who think the work speaks for itself (or, frankly, don’t have the communication chops Zohran does) this is how you combat the forces that want to make you fail

This is an incredibly important thing for the forces of good to understand. "Millions of citizens today find truth itself oppressive. They feel it oppresses their political agency. [...] Stop trying to lecture people back into truth. Instead, help them recover a sense of political agency."

1. Burn coal and oil. 2. Wait for Earth to heat up. 3. Grow olives on mount. 4. Lobby Christ to change venues. 5. Be new Israel. 6. Profit.

This is very real.

The single most annoying aspect of the Bluesky app is that it doesn't persist drafts. Put the phone down for a little too long and everything you typed is gone.

It's the Turning test.

I gotta say, "stop making RAM more expensive" is a much better swipe at AI dudes than "let's not let this slop replace the creativity of the human race".

This photo was taken in 1911 using glass plate technology by Herbert Ponting who was part of Scott's Antarctic expedition, The composition and detail are exquisite with the band of white snow/ice creating a perfect frame around the two people and the ship in the distance Iconic imo

"Culture fit" is a vague term that can easily collapse into affinity bias unless there's some rigorous process around it. Teams get weaker without a diversity of ideas and backgrounds. But I agree soft personality traits are as important as technical skill. Toxic "superstars" are a net negative.

I know it shouldn't, but watching this gives me vicarious back pain.

The biggest practical improvement in Opus 4.5 might actually be that it's fairly good at knowing when to stop and ask for advice, instead of careening off in some crazy direction. Chasing the agent back on track otherwise tends to eat up much of the productivity gains for me.

These delivery robots in China do not give a shit about anything

I love this thing. Imagine. Its late at night, an earthquake hits. You're trapped in your apartment building. Suddenly, outside the window: this robot baby with a glowing red smile and jet engines strapped to its back. "Follow me if you want to live," it says

As with many things, true mastery is knowing when the rules apply. "Code should be clean" and "it needs to be messy sometimes" are both true and both equally useless as maxims. The skilled developer knows when to tolerate the mess, and when not to, and there's no easy way to learn that.

The last few days did not move the main story forward AT ALL. It's just a guy eating little treats and doing nothing. And so many plot holes. He says, "no more little treats," and then he's eating another little treat in literally the next scene.

There's a caveat to this. In my experience, the leverage experts get from LLMs strongly depends on the type of project. For projects where you spend most of the time on analysis and a small fraction coding, the LLM has little to contribute. E.g. databases, filesystems, algorithms, network protocols.

I'm in favour of choosing wise independent voices by lottery.

Anecdotally, I find that current LLMs struggle with real world Inference to the Best Explanation situations. Things like diagnosing an illness, interpreting ambiguous T&C, predicting a food spoilage process; where there are many possible valid answers but one is more likely given the circumstances.

This is one of our favorites!

Really interesting research showing that while ultra-specialization in a single discipline might lead to better results early in one’s career, multi-discipline training and practice pays off big time in the long run. This applies to a range of professions from scientists to athletes and more

Progress is always 2 steps forward, 1 step back. But the negativity bias in the media will make it seem like that 1 step is far larger than the 2 forward

My bio mom is in the hospital right now for a COPD flare up and one of my spicy medical takes is that we have a lot research about how sleep deprivation is horrible for health and recovery but somehow this is totally ignored in clinical situations and waking you up every two hours is fine

The issue with doing things is that doing anything has a tendency to generate more things to do.

This is a great read, especially for those who claim to be able to detect “ChatGPT” written content - you’re most likely picking up on “Imperial British English” taught as a foreign language, and specifically the version taught to and used by Kenyans…

Hypothesis: Making all human knowledge immediately available to everyone only made us feel like we know more, while actually understanding just as little as people did in the past. The limiting factor was never the availability of information.