ANNOUNCEMENT: I wrote a book! It’s hard to talk about AI without a good frame exploring how it works, its history, and its future. So, I wrote a book on the topic. Order it here! Ideally in triplicate :-) https://t.co/5N3yPnT4F9

If it's such an excellent question surely you should've known to not do that

I think about this every time I'm in traffic. Point to point air travel would eliminate so much of the congestion and travel time we suffer through daily.

who has collated the best compendium of the types of AI errors, mistakes and reward hacks that we've seen happen in the last couple years?

If annoying users causes some to switch to a no-guardrail model, and this increases the chance that that model wins, what percentage probability should you put on "% annoy switch" and "% added win potential" before this is strictly EV negative? Solve for economics and safety.

Egypt’s name keeps surfacing, for a reason the news won’t tell you. Amanda Grace explains the prophetic connection.

Also, here's the DSL in pretty mermaid format.

Also on Android when I have a Pro conversation running, and I try to send an image to start a new conversation, it adds to the same one no matter what I do. I have to manually reselect another conevrsation and manually add it. Also, please fix!

. @openai friends. 5 Pro started defaulting back to Auto after one round in every conversation, so you have to re-select it. Extremely annoying, please do fix.

Glaze I agree, reluctantly, is useful. Coolocation can go jump in a well.

It’s time for Christian parents to realize that raising their children exactly like pagans does not produce young adults who cling to their faith.

The rest of the list is pretty bad too.

Do these dictionaries not understand what "word" means?

A very interesting twist on the housing theory of everything

a truly catholic thought-

Building Momentum: What Our Record Q3 Means for Abacus Shareholders | $ABL Chairman & CEO, Jay Jackson | @SubstackInc "I'm excited to share that Abacus has just delivered our tenth consecutive quarter of strong earnings growth—a milestone that reflects not just where we are

Not bad for a bit of work. Genuinely think there's a whole lot more we can do here. Will try write this up as a paper too at that point. Primer showed tiny low‑level tweaks can move the frontier, macro-evolution is even more fun!

Natural caveats apply ofc. The selection should be different/ more stringent. Longer runs needed to be sure. Ablation studies. Maybe worth doing a constant compute variant to compare to Primer paper. Have more baselines and mutations and benchmarks. Even so, pretty pleased.

In general I found the search over hunches approach to be super fun. I'm guessing this is popular inside frontier labs (I hope so), but also somewhat dumbfounded that this free alpha basically has been just laying around!

Oh here's the repo, and some insight as to what wins so far: https://t.co/3mXdVKbzR9

Why should startups use @Meow? We sat down with Steve Weidenbach, COO of @Blueprint_DeFi, to talk about it. Super proud to call them a customer. Meow is a financial technology company, not a bank. Banking services are provided by Grasshopper Bank, N.A; Member FDIC.

The "subway map" to see how core ideas are being composed into winning candiates. I somehow find this extremely pleasing to look at.!

One of the winners: - parallel attention+retention per layer - FiLM and lora modulation - windowed + quantized kv = rmsnorm and swiglu and small dim rope pretty consistent

Ok seems this doesnt exist so I built it. Transformer-evolution to find a better arch: - built a typed DSL to represent all transformer-ish architectures - evolutionary loop to mutate - selection to allocate compute to winners, multi objective fitness Some cool outcomes already-