Greg Medlock

Have any parents ever discovered that their 5 year old did indeed have a disease that made it too painful to walk?

We're searching for the next Director of the Duchossois Family Institute at UChicago The most exciting leadership role in microbiome science right now: lead faculty, cutting-edge cores, an FDA-approved facility for microbiome therapeutics Apply here: apply.interfolio.com/182203

@emmma_camp_ The methodology for the survey doesn’t seem to be public, and only 877 preschool teachers responded (far lower numbers for k-3). Amongst preschool teachers, I don’t think tenure is long enough for even a modest fraction of those teachers to have been teaching pre-COVID.

If it's parenting, why? What made parents just stop ensuring their kids develop basic motor skills?

Genuine question: Does anyone know why this happened? Most of these kids weren't even born until after the pandemic.

Hopefully this experiment shows other folks how peer review and collaborative work are going to evolve with AI... you can just "do the thing" now. I came away with a much better understanding and appreciation of GPU-kallisto! cc @lpachter

I definitely found that the bgzip vs. gzip point in @pmelsted's paper is critical. GPU-kallisto takes 10X as long (which I learned through trial and error, as I did not catch in Claude's implementation that it did not use bgzip). Opus 4.6 also was not terrific at figuring out compilation issues (though this is a pretty unique approach). If anyone knows of skills out there that optimize this, I'd love to learn more.

I was curious how @pmelsted's new GPU-kallisto implementation affects compute costs. Given the massive (50X!) improvement in speed, this wasn't straightforward. Instead of just asking "ackchyually, is it cheaper", my friend Claude and I did the analysis!

Excited to share this preprint that describes my latest work on using GPUs to accelerate processing of RNA-seq data. The title says it all: "RNA-seq analysis in seconds using GPUs" now on biorxiv biorxiv.org/content/10.648… Figure 1 shows they key result

@SynabunAI @strnr Luckily it was very easy to catch—Claude at least put that info up front in its summary. Now have a skill for networked, private projects…

@MedlockGreg @strnr Classic. It optimizes for "works" before it optimizes for "safe" unless you're explicit about it. Adding a security checklist to CLAUDE.md actually helps - once it's part of the default context it stops treating bind addresses as an afterthought.

"Shall I implement it?" gist.github.com/bretonium/291f…

@hhlee Locked into windows, WSL makes corporate machine unusable. You scp to view figures?

@MedlockGreg Terminal is all you need. iTerm for Mac?

Looking at VS Code alternatives for Claude Code to get full access to Claude Code features (e.g., no remote-control in VS code). What are the best terminal options? Needs to display locally on windows while sshing to EC2 where code is actually running, and support image/pdf viewing.

@pmelsted FYI, I decided this is easy enough with Claude that I should just do it! CPU benchmarking across a handful of instance types is done, should have the same for GPU in a day or two (rate limiter is GPU quota for my personal AWS account).

@pmelsted This is a super awesome contribution. But, the 50X improvement being cited is with a $400 CPU vs ($2k GPU + $400 CPU) config. Could you benchmark across some common EC2 instances for a cost-controlled speed comparison?

@GI_Pearls @tberzin Do doctors not know that their patients are not automatons?

@tberzin If docs knew their messages are rated on empathy, would they write differently? I feel like current LLMs value empathy over accuracy ( they certainly don’t offend). Docs( and many other humans in technical fields)value accuracy above empathy. No?

Physicians were rattled when patients rated Epic's AI messages as more empathetic than their own words. Apparently we are in good company: "NY Times readers favor AI writing over the work of Cormac McCarthy and Carl Sagan": nytimes.com/interactive/20…

@NEndoscopy @tberzin Empathy is not mutually exclusive with telling people things they do not want to hear, and I’d argue that you cannot be empathetic as a physician if you are not communicating complete information to a patient.

@tberzin Depends if we think this is the most important outcome though. Empathy is important but one of our roles is often to tell people things they don’t want to hear but is actually critical for their health…

@dshahh @cremieuxrecueil @RuxandraTeslo Academic would never do this outside of compassionate use IND though? In which case they cannot make money on it?

@cremieuxrecueil @RuxandraTeslo You can get much cheaper if you go academic GMP vs industry; but also depends on if you’re talking personalized or not; mRNA vs peptide vs cell tx, etc.

Personalized cancer vaccines cost $1.5–3M/ patient batch to manufacture today. Many experts say it should be ~$100K or less. Why the gap? Secretive CDMO pricing, limited competition & risk-averse pharma. Reporting on this. pls DM if you have on the ground experience. Need help!

@cremieuxrecueil @RuxandraTeslo Where can I learn more about obtaining these vaccines? Not in need, just curious how the hell this works…

@RuxandraTeslo For the custom cancer vaccines I've helped people get, the price is a lot less than $1.5-3m. Where's that cost estimate from?