Magic words sometimes work! If you ask for "advice" rather than "feedback," when trying to improve, everything is different. People are bad at evaluating others, but much better at giving advice, so you will get much more actionable & useful information. hbs.edu/ris/Publicatio…

Requiring pre-registration made a dramatic change in the reported outcomes of drug trials

One like = one spicy take about AI policy.

Just came across this weekly sprint update template from Y Combinator Feels like this sort of fast, informal accountability + reporting would be suuuper useful to a lot of people who have just started working on something new

"Language Models Can Teach Themselves to Program Better" This paper changed my thinking about what future langauge models will be good at, mostly in a really concerning way. Let's start with some context: [1/11]

Screenshots as a universal API

Building a startup (or working at a small company) can feel like hiking up a mountain alone, with no feedback and little feeling of progress. Some thoughts of how to handle the journey: hajak.se/growing-when-y…

We're still grappling with LLMs -- and multi-modal models will be a whole new universe.

AI influencers was the end-state Marvin Minsky had in mind

Natural language for handling the long tail of edge cases

Procrastination levels, ranked from least to most dangerous: 1. Do something fun to avoid the Big Task 2. Do a Little Task to avoid the Big Task on your to do 3. Prepare your tools for the Big Task 4. Upskill for the Big Task 5. Pretend the Big Task isn’t the Big Task

The number of people living in extreme poverty fell by ~100,000 yesterday. (This has been true every day since 1990, on average.)

Limiting belief issue

This is wild. H/T ImportAI

Excited to announce that Plymouth has achieved over 100 O1 and EB1 US work visa approvals for extraordinary individuals to build here in America.

in the short term, we’re rewarded for disconnecting from our bodies. but in the long term…

🚨New applied AI interpretability startup just dropped: Goodfire Eric Ho and Dan Balsam (founding team of RippleMatch) are joining forces with Tom McGrath (ex-GDM mech interp lead) and Nick to build the IDE for Software 2.0. The plan...

Distributed training will be a big deal if it scales beyond 1.2B models: 1. Electricity needs would no longer be the limiting constraint on big training runs. 2. Govts will struggle to monitor training runs. (This is only an early sign, but worth watching)

maybe no longer "when is the first billion-dollar training run?" but "when is the first 100k inference?"