wtf now my agents found the first bacteria to use new stop codons!!
After this TAG result, a chat w/ friend inspired the question: do any bacteria use other stop codons?
Overnight, my guys found myxobacteria that stop at TTA, normally read as an amino acid.
Evidence:
1) These bacteria lost the tRNA that normally reads TTA as leucine.
2) Their RF2, the protein that recognizes stop codons, has two changes: one in its codon-reading tip (SPF → SPY) and one nearby (L → F). No other species in GTDB has both.
3) We compared the same genes in these bacteria and their closest relatives. Where a relative's gene ends at a normal stop but the gene here keeps going, the very next codon is TTA 11% of the time (252 of 2,339 genes). In the relatives, it's 0.05% (1 of 2,025).
4) TTA swaps with the normal stop codons. Where one of these genomes ends a gene at TTA, another genome from the same group ends that gene at a normal stop 78% of the time.
Method:
Claude agents on my Mac mini scanned all 199,923 bacterial and archaeal genomes in GTDB for amino-acid codons unusually common near the ends of genes, where a hidden stop would sit. In 4,599 genomes, they then matched each gene to corresponding proteins and checked which codon sits exactly where the known protein ends.
Closest prior work:
TTA works as a stop in some mitochondria and plastids (Barcytė et al. 2025), and CGA was proposed as a stop in Asgard archaea (Melnykov 2026, PNAS), but none in bacteria until now.
One of these bacteria, 1-FT3.2, is kept alive in Finland (Demina et al. 2025). Possibly its proteins can be measured to confirm this.
I discovered the first bacteria that no longer use TAG as a stop codon.
Almost all life ends its genes with the same 3 "stop" signals: TAA, TAG and TGA.
Bacteria were known to reassign TGA since the 80s. Archaea reassigning TAG was published in Science last year.
🧬 1/n
Hey bio twitter, I was sick in bed this weekend but luckily my clankers were clanking
This time they found the first known case of genome-wide frameshifting in bacteria!
A whole order of gut bacteria skips a letter of its genes at CGG.
How they found it:
My agents aligned one genome from every bacterial and archaeal family in GTDB to known proteins, this time allowing for frameshifts. Every genome has a few apparent frameshifts (assembly errors, dying genes), spread thinly over all codons. In one TANB77 (uncultured Clostridia) genome they piled up at CGG.
So we tested every TANB77 genus, finding that:
- 54 of 64 genera show the skip, and all 54 have lost the CGG tRNA
- Related proteins have glycine at 1,237 of 1,458 skip sites, vs 6% at other frameshifts
- 59% of all +1 frameshifts in TANB77 fall exactly at CGG (1,026 of 1,744); the rest are the usual background noise, scattered thinly over all the other codons.
- Their phages skip CGG too: 58% of phage CGGs sit where a skip is needed, vs 2% in phages of other Clostridia
- It's not a sequencing error or RNA editing: DNA reads (140 of 140) and gut RNA (305 of 305 read pairs) keep the C
Closest prior work:
Genome-wide frameshifting is known in the ciliate Euplotes, and glass-sponge mitochondria skip a U the same way, but at only 2 sites.
@inperience That’s amazing! I hope that they enjoy hearing about it!
I am unsure about how I want to handle preprints. Since I’m not in the field, I don’t feel super comfortable writing it myself. Any thoughts on this?
@Gabogonzalez515 No, these results are just computational.
I checked and these bacteria haven’t been cultured before, but hopefully in the near future some collaboration can make this happen!
To give a bit more insight into why I searched for frameshifts:
In retrospect, both of my recent results (TAG reassignment and TTA as a stop: https://t.co/HTDUEQfbLr) came from searching for things that break the default assumptions of the genetic code and thus confuse existing search software. In these cases, irregular stop-codon behavior.
So a natural next question was: what other irregular behavior might be out there that existing software would miss by default? Frameshifts are already known in a few organisms, so they were a natural place to start.
Since then I've been trying to push this further, asking Fable to brainstorm exotic behaviors that might exist but have never been seen, then running genome searches for them. So far mostly negatives and a couple of leads, but I’ll keep you all posted.
Data and code: https://t.co/Jt8XUHHJup
wtf now my agents found the first bacteria to use new stop codons!!
After this TAG result, a chat w/ friend inspired the question: do any bacteria use other stop codons?
Overnight, my guys found myxobacteria that stop at TTA, normally read as an amino acid.
Evidence:
1) These bacteria lost the tRNA that normally reads TTA as leucine.
2) Their RF2, the protein that recognizes stop codons, has two changes: one in its codon-reading tip (SPF → SPY) and one nearby (L → F). No other species in GTDB has both.
3) We compared the same genes in these bacteria and their closest relatives. Where a relative's gene ends at a normal stop but the gene here keeps going, the very next codon is TTA 11% of the time (252 of 2,339 genes). In the relatives, it's 0.05% (1 of 2,025).
4) TTA swaps with the normal stop codons. Where one of these genomes ends a gene at TTA, another genome from the same group ends that gene at a normal stop 78% of the time.
Method:
Claude agents on my Mac mini scanned all 199,923 bacterial and archaeal genomes in GTDB for amino-acid codons unusually common near the ends of genes, where a hidden stop would sit. In 4,599 genomes, they then matched each gene to corresponding proteins and checked which codon sits exactly where the known protein ends.
Closest prior work:
TTA works as a stop in some mitochondria and plastids (Barcytė et al. 2025), and CGA was proposed as a stop in Asgard archaea (Melnykov 2026, PNAS), but none in bacteria until now.
One of these bacteria, 1-FT3.2, is kept alive in Finland (Demina et al. 2025). Possibly its proteins can be measured to confirm this.
Some random YC startup
- “We are the first ones to think of doing bitter lesson for trading” (there’s 4 more in the same YC batch)
- Please trust me bro, this has never been done
Meanwhile EquiLibre
- Literally has Rich Sutton as an angel investor and advisor.
- Founders left DeepMind to build this Neolab even before ChatGPT came out. Before that built the first AI to beat humans in no-limit poker.
- Already has highly profitable RL agents in the live market, trading billions of dollars every day.
wtf now my agents found the first bacteria to use new stop codons!!
After this TAG result, a chat w/ friend inspired the question: do any bacteria use other stop codons?
Overnight, my guys found myxobacteria that stop at TTA, normally read as an amino acid.
Evidence:
1) These bacteria lost the tRNA that normally reads TTA as leucine.
2) Their RF2, the protein that recognizes stop codons, has two changes: one in its codon-reading tip (SPF → SPY) and one nearby (L → F). No other species in GTDB has both.
3) We compared the same genes in these bacteria and their closest relatives. Where a relative's gene ends at a normal stop but the gene here keeps going, the very next codon is TTA 11% of the time (252 of 2,339 genes). In the relatives, it's 0.05% (1 of 2,025).
4) TTA swaps with the normal stop codons. Where one of these genomes ends a gene at TTA, another genome from the same group ends that gene at a normal stop 78% of the time.
Method:
Claude agents on my Mac mini scanned all 199,923 bacterial and archaeal genomes in GTDB for amino-acid codons unusually common near the ends of genes, where a hidden stop would sit. In 4,599 genomes, they then matched each gene to corresponding proteins and checked which codon sits exactly where the known protein ends.
Closest prior work:
TTA works as a stop in some mitochondria and plastids (Barcytė et al. 2025), and CGA was proposed as a stop in Asgard archaea (Melnykov 2026, PNAS), but none in bacteria until now.
One of these bacteria, 1-FT3.2, is kept alive in Finland (Demina et al. 2025). Possibly its proteins can be measured to confirm this.
I discovered the first bacteria that no longer use TAG as a stop codon.
Almost all life ends its genes with the same 3 "stop" signals: TAA, TAG and TGA.
Bacteria were known to reassign TGA since the 80s. Archaea reassigning TAG was published in Science last year.
🧬 1/n
I discovered the first bacteria that no longer use TAG as a stop codon.
Almost all life ends its genes with the same 3 "stop" signals: TAA, TAG and TGA.
Bacteria were known to reassign TGA since the 80s. Archaea reassigning TAG was published in Science last year.
🧬 1/n
@OmicsOmicsBlog Thank you for the kind words, that's the best kind of compliment :)
Wondering what you think the next steps should be, both for this research and who to chat with? I'm also trying to develop a better sense of which problems are actually interesting and which are just curiosities
@GrinsteinJ@leviathan77x Mac mini m6 32gb. I actually bought it 2 weeks ago to start on this project because before I was mostly just running on MacBook Air overnight.
hey will this is awesome! i've recently been working towards autoresearch outside of math/cs.
astronomy is amazing bc of how much public data there is.
here are some things that have turned up:
- JPL's newest orbit file for Uranus's small moons is way off. in JWST images, Ophelia is ~104,000 km from its predicted spot
- SOHO's sun-watching camera caught 2 known novae 71–81 days before ground sightings
- a baby star had a outburst in 1900, ~130× brighter than usual for 5 months. it was found recorded on old photo plates but never reported afaik
- 18 near-Earth asteroids on Harvard photo plates from 1898–1983, decades before discovery
- 3 deep pits on Mars missing from the USGS cave catalogue, found by searching 1,432 HiRISE images
- 12 untagged red sprites in 1.9M ISS astronaut photos
would be super cool to chat with your team to get some of these verified, and maybe share autoresearch setups/tricks
I have a major result to share.
I’ve found conclusive forensic evidence that the top 4 Chinese AI labs—DeepSeek, Moonshot, MiniMax and Ant Group—have built their newest models on NVIDIA’s most advanced Blackwell chips, which are barred from export to China.
This sounds insane, but I discovered that almost every AI model carries a hidden signature of the GPU that trained it. When a model is first set up, the starting values of its weights are laid out in a pattern that depends on the exact GPU. Each combination of GPU and seed number produces a unique pattern—a sort of fingerprint of the hardware used to create the model. Traces of that pattern survive training into the published model.
Again, this evidence isn’t based on any insider information or leaking. It’s based on a forensic analysis of the model weights released by the companies themselves.
I searched more than six trillion combinations of chip layout and seed number and matched them to open-weights models. In doing so, I found conclusive evidence that six Chinese AI models—including DeepSeek-V4-Pro, Kimi K3, and MiniMax-M3—were initialized on NVIDIA Blackwell systems. Four of the models match Grace Blackwell, the GB200 and GB300 parts in NVIDIA’s flagship liquid-cooled rack-scale systems.
I then verified these results on real hardware. Running unmodified PyTorch with the seeds I’d recovered, a Blackwell B300 reproduced DeepSeek-V4-Flash’s untouched starting weights bit for bit, and a rented GB200 reproduced Kimi K3’s to a correlation of 0.999995. A Hopper H100 reproduced none of DeepSeek-V4-Flash’s, and changing the seed by one broke the match. (DeepSeek’s seed is 953049583 by the way.) The method also was verified to match 24 known open-weights models to the hardware that developers said they used.
It is exceedingly unlikely that these results are a coincidence. Across six trillion combinations, the best accidental match should reach only about 8 sigma (i.e., eight standard deviations above what random chance would produce). The real matches range from 49 to 1,385 sigma, and even the weakest single test hits 20.
These were data-center-scale training runs. With some analysis of release schedules and model details, I estimate that the largest, Kimi K3 and DeepSeek-V4-Pro, used 2,000 to 3,000 Blackwell GPUs each. Clusters of this size cost on the order of $150 to $200 million each to buy, and across these labs alone would add up to more than half a billion dollars of export-controlled GPUs, not counting experimental compute or otherwise. It’s quite likely that we see multi-billion dollar expenditure on export-controlled GPUs.
I am highly confident in this result, and it has been independently verified by senior researchers at OpenAI and Anthropic. You can read the full report at https://t.co/YOtzND0Jqk.
There is much more I’ve identified about these models through other forms of checkpoint forensics like this, and will have more to share soon.