
🩺 doctor defected to tech 👨🏼💻 prev engineer @exolabs | medicine @uniofoxford | ML @imperialcollege
Joined August 2021
- Tweets215
- Following807
- Followers8.5K
- Likes667
Seth Howes retweeted
I lost my dog because we could not build her next mRNA cancer vaccine fast enough
Today we are announcing a $4M seed led by @foundersfund to change that.
Curing dog cancer
We are bringing personalized mRNA cancer vaccines to dogs and building the evidence to show how well they work.
We have raised a $4M seed led by @foundersfund to make it happen
This investment allows us to launch a clinical trial to validate the personalized cancer treatment protocol we developed for Rosie.
If your dog has cancer or you know someone whose dog does, please tell them about Gamgee.
Seth Howes retweeted
My dog Rosie was given months to live. Chemo and immunotherapy failed
I used chatgpt + computational genomics to design a personalised mRNA cancer vaccine for her tumour. Several of her tumours shrank
Today we launch Gamgee (YC S26): personalised mRNA cancer vaccines for dogs
Seth Howes retweeted
New Huang lab publication at Nature. Our newly developed technique: transient Pore Analyte Looping (tPAL) now achieves “chop and measure” nanopore sequencing of peptide, in an amino acid by amino acid manner. nature.com/articles/s41586-0…
Seth Howes retweeted
Big news today. Science's PRIMA visual prosthesis, our retinal implant for macular degeneration, has received its CE mark, which means marketing approval in Europe. This is a huge milestone to begin to have a device like this commercially available to regular patients. 1/
this is insanely cool. running a language model on the same microprocessor that powered the apple ii in 1975.
I built a 50k parameter language model to run on an 8-bit 6502 processor from 1975. It learnt to spell words and construct basic grammar!
The model is build on a Mamba backbone (attention’s KV cache not fit in 32KB of memory) with a character-level tokenizer.
The MOS-6502 doesn’t have a ‘multiply’ instruction - only 8-bit integer addition. I therefore used ternary weights to reduce matrix multiplication to a series of add/subtracts operations.
Technical write-up here: mattbeton.com/blog/bitnet-65…
Source code: github.com/MattBeton/BitNet6…
Seth Howes retweeted
I built a 50k parameter language model to run on an 8-bit 6502 processor from 1975. It learnt to spell words and construct basic grammar!
The model is build on a Mamba backbone (attention’s KV cache not fit in 32KB of memory) with a character-level tokenizer.
The MOS-6502 doesn’t have a ‘multiply’ instruction - only 8-bit integer addition. I therefore used ternary weights to reduce matrix multiplication to a series of add/subtracts operations.
Technical write-up here: mattbeton.com/blog/bitnet-65…
Source code: github.com/MattBeton/BitNet6…
Replying to @NikoMcCarty
This is crazy useful. I’ve found Chinese tech news to be inaccessible to westerners.
For example, many leading Chinese tech distribute content on WeChat through the “Official Accounts” feed.
I have a web scraping cron job and translation pipeline so I can aggregate this info into a single feed
Replying to @VishalGulati_
Thankfully Brad, and the minority of people who do novel things, don’t fit into the bucket of “most smart people”
Anyone can now sequence their genome at home.
i have now sequenced my own DNA 5x at home and learned so much about myself
i wrote out the protocol here so anyone can follow it and talk to their DNA: bradleywoolf.com/links-1/seq…
a lot of people helped with this, they are all mentioned below
Seth Howes retweeted
Automated liquid handlers for biology are much older than I thought. A rough timeline:
- The 96-well plate is invented in 1951 by a Hungarian physician, Gyula Takátsy, who physically drilled each well out of a plate of Lucite, a type of acrylic.
- John Sever made the first "modern-looking" 96-well plate by punching it directly out of a sheet of plastic.
- The first automated liquid-handling device was released in 1967 (!!!). It was called the Autotiter, and was developed by a man named Tom Astle. I found a few sentences on Astle's company:
"Tomtec has succeeded since 1967 by working with its clients to help solve their liquid handling needs. From 1967 to 1981, the Company had essentially one employee, Tom Astle. From 1967 to 1971, the Company operated as Astec, Inc. In 1971 the Company's name was changed to Tomtec, and incorporated. Tom Astle has been the President and CEO since that time. In 1967, Tomtec developed the Autotiter. This automated what were then the manual microtiter techniques. The primary market was serology and virology. The original instruments were an essential element of Smith Kline's rubella vaccine program." (Photos of the original patent filing below.)
- In 1986, Beckman gets into the microplate automation space. They release the Biomek 1000 and mark it as a "fast and accurate...robotic workstation" that sits on a scientist's bench, "takes over tedious manual tasks and integrates the work of four different instruments." I found these details in a Beckman ad, which was published in a 1987 issue of the "Journal of Automatic Chemistry." Photos below.
The Beckman device looks fully modern; it could even connect to an IBM computer so scientists could program new methods on it. I couldn't find a launch price, but the device was similar to the Zymark Zymate, and that device cost about $35,000 when it released in the 1980s, which would be equivalent to hundreds of thousands of dollars today.
Seth Howes retweeted
The pMHC binding with an antibody is so crazy to me. If the binding is general across patient MHCs, you can basically target cancer driver mutations with antibodies (which Revolution meds showed was effective towards pancreatic cancer). Step change, congrats to @nablabio team!
Today, we expand zero-shot drug design beyond binding to the design of multifunctional medicines, the intracellular proteome, and state-of-the-art atomic precision with our model, JAM-2.
In a new report (below), we show:
1. The first drug-grade, fully computationally designed multispecific antibodies against five peptide-MHCs: Routine picomolar T-cell activation/cell-killing EC50s, >100-fold selectivity, and drug-like developability
2. The first fully generatively designed, drug-grade dual-variant KRAS G12 multispecifics: They recruit primary T-cells from human donors to kill G12V and G12C presenting cells at pM to single-digit-nM potency, completely sparing wild-type.
3. Atomic accuracy, from sequence alone: Angstrom-level agreement between Cryo-EM and JAM-2 de novo designs, requiring only target sequences (not structure) as input.
4. Unrivaled speed with an AI-native in-house wet lab: Designed, built, and tested five programs in one parallelized campaign, end-to-end in-house in ~6 weeks.
5. A higher validation bar for AI-generated drug candidates: In a field increasingly rife with hype and uneven standards of proof, we provide the highest quality public wet-lab validation of AI-designed antibodies to date. We share experimental methods in full, and invite folks to adopt and build on these standards.
Truly individualized therapies will be the most important contribution of AI in drug design. These advances help accelerate this future.
Replying to @Sephirsloth
I was taken by a brilliant synthetic biologist to this quaint cafe today.
Funny thing. I must have missed the ‘dishonest hordes’. Where were they sitting?
Replying to @KevinJDCS
Everyone I met today was in their 20s, intelligent, and wanted to achieve greatness.
All lovely people as well.
There’s a cafe in Shenzen’s biotech hub that gives you free coffee for a month if you publish in Nature or Science.
Replying to @LeroyLeroi001 @Jansen0781
This is empirical. Walk down any street in Shenzen and look at the variety of premium Chinese brand EVs being driven.
These cars aren’t exported: they don’t exist in the US.
Replying to @IgRecon
I’m not making a value judgement, or a broader comment on the cost paid to get here.
I’m simply observing that most Chinese have access to higher quality, cheaper consumer technology relative to US consumers.



















