Craig Tashman

"The interviewer, @DouthatNYT, broke down crying during the conversation. And when he did, Sasse laughed. Not unkindly, but the way a man laughs when the heaviness and the lightness of something are both true at the same time and he’s decided not to pretend otherwise."

What could go wrong? anthropic.com/research/small…

So basically we can’t trust LLMs anymore?

Silly to pay people for commodity thought work, pay people for humanness itself

@gilrdb @zotero DM will be hard, but the form on the website will route to the right people. The team will be back online on Monday. I’m sorry about this, hopefully we can find out why you’re seeing an issue here quickly!

@CraigTashman @zotero i did submit through the contact form on the liquidtext website. thanks for the quick reply! when can i expect an answer? any chance we can DM?

.@CraigTashman there is an issue with pdf import using the @zotero integration that repeatedly crashes the liquidtext on iPad. this is unfortunately not usable at its current state. assistance needed

@andy_matuschak Thanks Andy! I’m excited to see where you take this, anyone can pull off the vision for this bazaar it’s you

@andy_matuschak It’s a compelling vision, Andy! But I wonder if enough designers, human or AI, will be able to design consistent and cohesive experiences without the structure of modularity?

⭐ New talk! andymatuschak.org/tat Coding agents might help us finally break out of two cages: the app model, which traps computing in one-size-fits-all silos; and programming as a specialization, which has crowded out cultures of imagination and domain insight.

This is big... Anthropic just announced a model so powerful they won't release it to the public out of fear over the damage it will cause 😨 Claude Mythos Preview found thousands of zero-day exploits in every major operating system and web browser... The numbers are hard to believe: > $50 to find a 27-year-old bug in OpenBSD, one of the most security-hardened operating systems ever built > Under $1,000 to find AND build a fully working remote code execution exploit on FreeBSD that grants unauthenticated root access from anywhere on the internet > Under $2,000 to chain together multiple Linux kernel vulnerabilities into a complete privilege escalation exploit For context: these are the kinds of findings that previously required elite security researchers working for weeks. Anthropic engineers with no formal security training asked Mythos to find exploits overnight. They woke up to working code the next morning. The results were so impressive Anthropic assembled Apple, Google, Microsoft, Amazon, NVIDIA, and seven other organizations into Project Glasswing: A $100M defensive coalition. They're not releasing this model publicly. Instead, they're racing to patch the world's infrastructure before models like this proliferate.

CS Lewis actually agrees. He argues that God effectively gives all of us this chance, but whether we “mature” into glorious creatures grounded in love, or hateful vain shadows grounded in nothing but ourselves, comes from the choices we make now. See “The Great Divorce.”

@dunkhippo33 This will become increasingly true

This is what most people don't realize - you can actually build big businesses without VC dollars. Moreover, a lot of the big raises happen BECAUSE the company didn't need money:

LLMs are amazing, and have sped up some of our engineering by 10x. But if you think these things are anywhere close to human thought, you need to talk to more humans!

@fchollet How far do you think we are from fluid intelligent AI? Do you think LLMs will be able to get us there?

When high-fluid intelligence systems start to show up, they will immediately take over the knowledge-dependent ones. Because they will be able to scale their knowledge just as well as legacy systems (knowledge gathering is the easy part), while their ability to recombine and apply that knowledge will create a performance gap that preparation alone will never bridge

People struggle to differentiate fluid intelligence from knowledge because, given enough preparation, memorized templates become a solid substitute for on-the-fly adaptation

Terence Tao put it plainly: there is no evidence that LLMs exhibit genuine creativity. Yes, they have solved some Erdős problems. But these are low-hanging fruit, questions that attracted little attention and that yield once the right existing techniques are applied. That is not creativity. That is search plus recombination. Yes, LLM outputs can look impressive. But look at who is impressed: typically non-experts. Experts know very well that LLM performance gets terrible when you approach the frontier of human knowledge. And this is not a temporary gap. It reflects a structural limitation. We do not fully understand human creativity. But we do know a key property: Conceptual leaps: the ability to generate new representations, not just recombine existing ones. LLMs do not do this. They interpolate in representation space. They operate within existing conceptual frameworks; they do not create new ones. This is why we haven’t “yet seen them take the next step”.

LLMs are a distillation of human thought as it relates to the subject matter in the training set. I don’t see any reason to think they can replicate human thought more generally than that?

Sure, LLMs are nowhere close to the limits of human thought. But my fear is that LLMs are so good at doing our day to day tasks that we’ll stop doing them, stop building the deep domain expertise where our unique cognitive abilities can shine, and end up in a cognitive winter for a few generations.

Using Skills well is a skill issue. I didn't quite realize how much until I wrote this, the best can completely transform how your team works.

Good post on the incentive misalignment between AI labs and customers, and how software makers can take advantage of it. Worth a read.

New research from Tsinghua, Peking University and other top labs taught a humanoid robot to play tennis using scattered human movement clips instead of perfect match data. The big deal here is how the team solved the data problem for physical robots. Usually, teaching a robot to do something highly athletic like playing tennis requires perfect, continuous tracking data of professional human players. Getting that kind of flawless 3D physical data during a high-speed match is extremely difficult and expensive. This paper bypasses that massive hurdle entirely. Instead of needing perfect full-match data, the researchers just used short, disconnected, and imperfect clips of basic human swings. The AI system uses these rough clips as a basic hint for how a swing should look, and then a physics simulator corrects the physical errors so the robot does not fall over while swinging to hit the ball. Because they proved they can take messy, fragmented human data and turn it into a smooth, highly dynamic robot athlete, this means we can start teaching robots all sorts of complex physical tasks without needing to record perfect human demonstrations first. It severely lowers the barrier to making robots useful in fast, unpredictable physical environments. The robot successfully tracked fast incoming balls and consistently hit them back to specific target zones while looking surprisingly natural.