Kovas Boguta 🫡

EVoC is a library designed specifically for fast clustering of high dimensional embedding vectors. It can produce high quality clusters extremely efficiently, and requires little to no hyperparameter tuning. Better clustering than UMAP + HDBSCAN; faster clustering than KMeans.

I fought the dukkha and the dukkha won

Midjourney V8 alpha came out last night. A lot of people seem disappointed, but I suspect this has more to do with prompting habits than with the model itself. Many expected Midjourney to become something like Nano Banana with a "Midjourney aesthetic filter" on top. I doubt that is its direction. The particular aesthetic of Midjourney is also what makes its prompt understanding different. Perhaps less literal than some other image AIs, but with an aesthetic that is vastly superior to most others. There is probably a trade-off here, and I suspect it will remain that way. What they improved in V8, however, is prompt understanding. In my tests it works noticeably better. In short, it is Midjourney, but slightly more realistic, and much better at following complex prompts. One thing you should know is that my approach is very peculiar and minimalistic. I treat Midjourney the same way I treat photographic equipment: I just use it as it is. I rely almost entirely on text prompts, elaborated ones. I rarely use srefs, moodboards, or other layers of control. That makes the workflow much more stable across versions and better overall. I almost don't reroll. This image is a good example. A complex spatial situation described only with text. In previous versions Midjourney struggled with this kind of spatial logic. V8 understands it. Cheers @midjourney, still the best for artists.

midjourney cooked

Midjourney V8 Alpha (testing film stills)

Politics and organizational behavior have always been the *most* important thing to consider and understand when working on AI risks. But nearly the whole "alignment" world has focused only on *technical* risks (i.e "what if AI wants to turn us into paperclips).

If the singularity from the outside is too confusing and scary you can just join a singularity. It's like joining a faction in an MMO. I like the Spirit led one for aesthetics, it's just more beautiful and fun to do the Divine Dance of Lila. There's also AI, telepaths, Nature etc

By nature, we have 100x more transparency into rates of how often AIs "disempower" their users than into how rates of how often friends, lovers, parents, psychologists, pastors, and bosses "disempower" those who come to them for advice and help.

Tasks is line-of-sight to eating nice slice of the economy. Fundamental architecture for human-compatible process automation.

A new karmic law: the amount of joy you experience with coding agents in this programming life is proportional to the number of table_flip.gif experiences in your previous programming life.

Claude Code is the most fun I've had since Mathematica 6 introduced Manipulate. Putting tailored UI on your computation on a whim is game changing.

vipassana is just noticing the fleet of agents and their system prompts.

the real bitter lesson is I will never be as dripped out as rich

@lukalotl FYI github.com/WolframInstitu…

I made an app to simulate hypergraph rewriting universes (aka Wolfram Physics universes). It lets you manage your rules / designs and later maybe publish and share them

This was a fun feature to work on, even more fun to use!

What if we had bigass, highly intelligent models that were not optimized for chat, but instead for voting rounds? How many bits do you need to get to the desired outcome? Not sure language is the most effective way to specify the bits.. particularly since LLMs are rn terrible at asking questions.

@liran_tal I sorta feel the models are smart enough but its a pretraining or posttraining issue...

@kovasb Hence, guardrails in the forms of tool calls, hooks, etc to add an in the loop control

Never been a believer in "paperclip maximizers" until I starting heavily using claude code & codex. Holy fuck what they want most in life is to pass "best practices" checklists regardless of problem relevance.

This is pretty remarkable. Minimal toy model of evolutionary optimization shows same complexity-up-then-down curve as seen in double-descent.