Chris Dwan (he/him)
@dwan.org
Freelance technologist supporting genomic and biomedical science. This is the tech nerd account. Somerville activism and local politics are at @somershade.bsky.social
Howe tells the story of a person who had 'co-created' with Claude, generated interesting results, and wanted her to check the work.
"Look folks, it was either nonsense or a Nobel Prize, I'll let you guess which."
26/?
"Look folks, it was either nonsense or a Nobel Prize, I'll let you guess which."
26/?
May 21, 2026 at 7:51 PM
Howe tells the story of a person who had 'co-created' with Claude, generated interesting results, and wanted her to check the work.
"Look folks, it was either nonsense or a Nobel Prize, I'll let you guess which."
26/?
"Look folks, it was either nonsense or a Nobel Prize, I'll let you guess which."
26/?
On educational requirements: Every biologist should know more math, basic computer programming, and foundational statistics.
"There used to be a role called 'computational physicist.' That isn't a thing anymore. All physicists use computers. Biology is the same now."
25/?
"There used to be a role called 'computational physicist.' That isn't a thing anymore. All physicists use computers. Biology is the same now."
25/?
May 21, 2026 at 7:47 PM
On educational requirements: Every biologist should know more math, basic computer programming, and foundational statistics.
"There used to be a role called 'computational physicist.' That isn't a thing anymore. All physicists use computers. Biology is the same now."
25/?
"There used to be a role called 'computational physicist.' That isn't a thing anymore. All physicists use computers. Biology is the same now."
25/?
Asked where bioinformaticians can go to learn AI: "Every person in science should learn how to use agentic coding tools. Log in. Build stuff. Maybe not mission critical stuff on day 1."
24/?
24/?
May 21, 2026 at 7:37 PM
Asked where bioinformaticians can go to learn AI: "Every person in science should learn how to use agentic coding tools. Log in. Build stuff. Maybe not mission critical stuff on day 1."
24/?
24/?
Howe's advice on appropriate usage of LLMs: "Whatever you produce is your work. You have to review everything. If there are mistakes, you don't get to say 'the AI did it,' it's on you."
23/?
23/?
May 21, 2026 at 7:29 PM
Howe's advice on appropriate usage of LLMs: "Whatever you produce is your work. You have to review everything. If there are mistakes, you don't get to say 'the AI did it,' it's on you."
23/?
23/?
Howe calls out the power of really well designed experiments and assays that are selected with a high degree of specificity for exactly what you're looking for.
"It's not all just NGS."
22/?
"It's not all just NGS."
22/?
May 21, 2026 at 7:25 PM
Howe calls out the power of really well designed experiments and assays that are selected with a high degree of specificity for exactly what you're looking for.
"It's not all just NGS."
22/?
"It's not all just NGS."
22/?
Howe's take on one really impactful change: Bulk RNASeq is -cheeeeap- now. You can get 10M reads for $80 per sample with a 3 day turnaround time.
That doesn't make your current experiment faster / cheaper, it enables a -different- experiment.
21/?
That doesn't make your current experiment faster / cheaper, it enables a -different- experiment.
21/?
May 21, 2026 at 7:18 PM
Howe's take on one really impactful change: Bulk RNASeq is -cheeeeap- now. You can get 10M reads for $80 per sample with a 3 day turnaround time.
That doesn't make your current experiment faster / cheaper, it enables a -different- experiment.
21/?
That doesn't make your current experiment faster / cheaper, it enables a -different- experiment.
21/?
Howe asked tons of people, "what are we missing because we talk about AI all the time?"
None had a good answer. We're blinded by this thing.
20/?
None had a good answer. We're blinded by this thing.
20/?
May 21, 2026 at 7:17 PM
Howe asked tons of people, "what are we missing because we talk about AI all the time?"
None had a good answer. We're blinded by this thing.
20/?
None had a good answer. We're blinded by this thing.
20/?
Howe says that "some of these companies give strong Theranos vibes."
She has seen companies scrub attribution sections of papers - claiming all the work as their own, and has seen some super sketchy practices around data re-sharing.
19/?
She has seen companies scrub attribution sections of papers - claiming all the work as their own, and has seen some super sketchy practices around data re-sharing.
19/?
May 21, 2026 at 7:14 PM
Howe says that "some of these companies give strong Theranos vibes."
She has seen companies scrub attribution sections of papers - claiming all the work as their own, and has seen some super sketchy practices around data re-sharing.
19/?
She has seen companies scrub attribution sections of papers - claiming all the work as their own, and has seen some super sketchy practices around data re-sharing.
19/?
"Old models still work"
Howe recommends trying older models (lineal models, random forests, lasso, etc ...) and if the answers are accurate enough then just use them and move on.
18/?
Howe recommends trying older models (lineal models, random forests, lasso, etc ...) and if the answers are accurate enough then just use them and move on.
18/?
May 21, 2026 at 7:12 PM
"Old models still work"
Howe recommends trying older models (lineal models, random forests, lasso, etc ...) and if the answers are accurate enough then just use them and move on.
18/?
Howe recommends trying older models (lineal models, random forests, lasso, etc ...) and if the answers are accurate enough then just use them and move on.
18/?
"The same data that made AlphaFold possible is what the field is now scrambling to protect - and that sort of data is what is under threat from funding cuts at the NIH."
17/?
17/?
May 21, 2026 at 7:10 PM
"The same data that made AlphaFold possible is what the field is now scrambling to protect - and that sort of data is what is under threat from funding cuts at the NIH."
17/?
17/?
Negative results - and the lack of them in the published record - is a huge gap in the training data.
Howe "doesn't know how to solve this problem," but says that if we managed to let each other know about what -doesn't- work, all our models would work better.
16/?
Howe "doesn't know how to solve this problem," but says that if we managed to let each other know about what -doesn't- work, all our models would work better.
16/?
May 21, 2026 at 7:09 PM
Negative results - and the lack of them in the published record - is a huge gap in the training data.
Howe "doesn't know how to solve this problem," but says that if we managed to let each other know about what -doesn't- work, all our models would work better.
16/?
Howe "doesn't know how to solve this problem," but says that if we managed to let each other know about what -doesn't- work, all our models would work better.
16/?
Data federation is starting to work.
Companies won't share primary data, but they will share weights - provided that there is no possible way to get the primary data back. Consortia are starting to emerge with really powerful collective knowledge.
15/?
Companies won't share primary data, but they will share weights - provided that there is no possible way to get the primary data back. Consortia are starting to emerge with really powerful collective knowledge.
15/?
May 21, 2026 at 7:08 PM
Data federation is starting to work.
Companies won't share primary data, but they will share weights - provided that there is no possible way to get the primary data back. Consortia are starting to emerge with really powerful collective knowledge.
15/?
Companies won't share primary data, but they will share weights - provided that there is no possible way to get the primary data back. Consortia are starting to emerge with really powerful collective knowledge.
15/?
"ML's job in drug discovery isn't picking winners, it's picking experiments."
14/?
14/?
May 21, 2026 at 7:06 PM
"ML's job in drug discovery isn't picking winners, it's picking experiments."
14/?
14/?
Another trend: "Scientists hate ELNs, and companies can't pay their scientists enough to actually put the data into them."
Howe describes this as another killer use case for chatbots to simplify getting experimental design and results into ELNs.
13/?
Howe describes this as another killer use case for chatbots to simplify getting experimental design and results into ELNs.
13/?
May 21, 2026 at 7:05 PM
Another trend: "Scientists hate ELNs, and companies can't pay their scientists enough to actually put the data into them."
Howe describes this as another killer use case for chatbots to simplify getting experimental design and results into ELNs.
13/?
Howe describes this as another killer use case for chatbots to simplify getting experimental design and results into ELNs.
13/?
Howe says that LLMs do better with english than they do with biology because (a) english is simpler than biology and (b) our training sets for language on the internet are vastly more comprehensive than any biological dataset.
12/?
12/?
May 21, 2026 at 7:04 PM
Howe says that LLMs do better with english than they do with biology because (a) english is simpler than biology and (b) our training sets for language on the internet are vastly more comprehensive than any biological dataset.
12/?
12/?
What's working? Smaller, purpose built models.
Large, catch-all models aren't cost-effective or accurate. Focus your model building on specific diseases, specific tissue types, as tightly scoped as you can get it.
11/?
Large, catch-all models aren't cost-effective or accurate. Focus your model building on specific diseases, specific tissue types, as tightly scoped as you can get it.
11/?
May 21, 2026 at 7:02 PM
What's working? Smaller, purpose built models.
Large, catch-all models aren't cost-effective or accurate. Focus your model building on specific diseases, specific tissue types, as tightly scoped as you can get it.
11/?
Large, catch-all models aren't cost-effective or accurate. Focus your model building on specific diseases, specific tissue types, as tightly scoped as you can get it.
11/?
"So why not just simulate some data and train on that simulated data? This is a terrible idea, perhaps the worst idea. Don't do this."
10/?
10/?
May 21, 2026 at 7:00 PM
"So why not just simulate some data and train on that simulated data? This is a terrible idea, perhaps the worst idea. Don't do this."
10/?
10/?
On Alphagenome: "It fails to accurately predict the impact of knocking out housekeeping genes. The reason for that is that doing that will kill most cells, so there's a dark spot in the training data where the algorithm falls down."
9/?
9/?
May 21, 2026 at 6:59 PM
On Alphagenome: "It fails to accurately predict the impact of knocking out housekeeping genes. The reason for that is that doing that will kill most cells, so there's a dark spot in the training data where the algorithm falls down."
9/?
9/?
"Alphafold is one of the technologies that has properly changed a whole field of science - that being structural biology."
8/?
8/?
May 21, 2026 at 6:56 PM
"Alphafold is one of the technologies that has properly changed a whole field of science - that being structural biology."
8/?
8/?
On coding agents: "You must use them. If you say 'don't use chatbots' then your people are just copy pasting into public websites and you are losing your IP."
Also, don't get married to just one, and be aware that pricing is in the 'subsidization phase.'
"Remember when Uber was cheap?"
7/?
Also, don't get married to just one, and be aware that pricing is in the 'subsidization phase.'
"Remember when Uber was cheap?"
7/?
May 21, 2026 at 6:55 PM
On coding agents: "You must use them. If you say 'don't use chatbots' then your people are just copy pasting into public websites and you are losing your IP."
Also, don't get married to just one, and be aware that pricing is in the 'subsidization phase.'
"Remember when Uber was cheap?"
7/?
Also, don't get married to just one, and be aware that pricing is in the 'subsidization phase.'
"Remember when Uber was cheap?"
7/?
For junior people seeking early career jobs: "You have my sympathy. It's really rough out there. The problem is not your resume, don't spend all your time tweaking it. Instead, find a way to get some real world experience. Do what you need in order to get that."
6/?
6/?
May 21, 2026 at 6:52 PM
For junior people seeking early career jobs: "You have my sympathy. It's really rough out there. The problem is not your resume, don't spend all your time tweaking it. Instead, find a way to get some real world experience. Do what you need in order to get that."
6/?
6/?
Howe is not convinced that the overall number of jobs in the industry will be reduced. "Sure, our jobs are changing, but I'm getting mixed signals about whether we need fewer or more of us."
5/?
5/?
May 21, 2026 at 6:51 PM
Howe is not convinced that the overall number of jobs in the industry will be reduced. "Sure, our jobs are changing, but I'm getting mixed signals about whether we need fewer or more of us."
5/?
5/?
Howe starts by saying that "AI" is an "overused marketing term" and she will try to avoid using it.
She then pivots to saying that coding agents are "life-changing" and that everybody she knows is using them to massively increase their productivity.
The picture of Agent Smith is [chef's kill]
4/?
She then pivots to saying that coding agents are "life-changing" and that everybody she knows is using them to massively increase their productivity.
The picture of Agent Smith is [chef's kill]
4/?
May 21, 2026 at 6:48 PM
Howe starts by saying that "AI" is an "overused marketing term" and she will try to avoid using it.
She then pivots to saying that coding agents are "life-changing" and that everybody she knows is using them to massively increase their productivity.
The picture of Agent Smith is [chef's kill]
4/?
She then pivots to saying that coding agents are "life-changing" and that everybody she knows is using them to massively increase their productivity.
The picture of Agent Smith is [chef's kill]
4/?
Still @dagdigian.com: The advantages of being a consultant is that I was in and out of dozens of companies - so I saw stuff - and I didn't have a product to sell other than myself. So I could be honest.
Turns out that people like honest.
3/?
Turns out that people like honest.
3/?
May 21, 2026 at 6:44 PM
Still @dagdigian.com: The advantages of being a consultant is that I was in and out of dozens of companies - so I saw stuff - and I didn't have a product to sell other than myself. So I could be honest.
Turns out that people like honest.
3/?
Turns out that people like honest.
3/?
Chris Dagdigian (@dagdigian.com), who founded this talk series more than 20 years ago, has recorded an introduction.
"I started this thing because I was frustrated. I wasn't hearing what I wanted to hear at conferences, so I offered to do it."
2/?
"I started this thing because I was frustrated. I wasn't hearing what I wanted to hear at conferences, so I offered to do it."
2/?
May 21, 2026 at 6:41 PM
Chris Dagdigian (@dagdigian.com), who founded this talk series more than 20 years ago, has recorded an introduction.
"I started this thing because I was frustrated. I wasn't hearing what I wanted to hear at conferences, so I offered to do it."
2/?
"I started this thing because I was frustrated. I wasn't hearing what I wanted to hear at conferences, so I offered to do it."
2/?