Pinned Tweet
Every engineer needs to hear the story of the Gossamer Condor at least once. Decades ago there was an early version of the X Prize: a cash reward for the first human-powered machine that could fly a figure-8 around two markers half a mile apart. A ton of serious teams went after it. They'd spend a year building one beautiful, carefully engineered machine, fly it, crash it, and spend another year rebuilding. This went on for almost two decades.
The team that won didn't try to build a perfect machine. They built a system for learning fast — something they could build, fly, crash, and repair the same day, then try again. They made it out of aluminum tubing, thin Mylar, piano wire, and tape. Cheap and flimsy, but every crash taught them something, and because they could crash and fix it multiple times a day, they out-learned everyone and won.
New concept the natural LLM an inverse mechanical Turk. A human pretending to be a LLM to get over organizational hurdles… everyone is used to getting advise from a LLM, getting people to accept feedback from a human is harder
Infinite amount of alpha in going over papers in microbiology published in the 60s 70s and re-evaluating them with fresh eyes/tokens
Random thought: is the closure of the BioWare game studio driven by model refusal in everything that mentions the word bio??? maybe it is just easier to switch to a different studio name than fight the rejections
If the best economical use case for your frontier intelligence is selling it by the token, you don't have frontier intelligence
I wonder how many mathematicians got effectively released from the task their grant assigned to them, and whether the last week will let us get a glimpse of what they were really interested in but thought would never get them a grant
The fact that protein embedding models understand sequence/structure/function dependency well enough that they can create synthetic homologs using iterative masked diffusion that can be used for MSA and improve protein folding prediction is nothing short of amazing
Random idea: You can use synthetic homologs for proteins using masked diffusion and create MSA from them. I was working with the ART proteins and needed better structural predictions so this got invented (will probably remove the deepmind protein SynthId) github.com/avi/dplm-syntheti…
I wonder how jagged is human intelligence without our noticing it… LLMs are so jagged they should come with a warning sign “beware sharp edges” but if someone evaluated humans without our prior expectations they would probably say the same about us
Are we really sure putting agents in a sandbox is a good idea??? A box filled with sand is exactly what they need to create more silicon, more chips and more agents
Random thought of the day… would a model trained on predicting the embedding of the next/prev protein in the genome be able to understand genomes and proteins better than others?
The plot thickens increase transcription errors in the ARTs, especially in the start of the array… I wonder what mechanism does that… includes the conserved repeats so probably not a culture diversity thingy
RNA based phages have a high mutation rates which means they can adapt to the host super fast but that limits their gene size to 4-5 genes, DNA based phages can be a lot bigger but because they have a low mutation rate, but that means that normally they have low adaption rates… ARTs solve the tension and create a locus that is super adaptable inside a stable genome
Continuing the ARTs saga, ARTs RNA is trans-splicing with the host RNA. ncbi.nlm.nih.gov/bioproject/… so we have the mechanism of how it helps the phage... now that I have precise cutoffs, openfold-3/boltz will be able to find the rest of the loop