they weren't kidding about claude fable 5 blocking biology: 38/40 questions refused on my private eval!

Nice to see Claude Opus 4.8 refusing fewer biology questions than 4.7. It's now also recognizing the error-prone nature of correctly reproducing long sequences: "I can't provide the exact full amino acid sequence of... as I'm not confident I can reproduce it with the precision required. Sequence data like this is the kind of information where small errors (a single wrong residue) can be significant, and I don't want to give you something that looks authoritative but contains mistakes."

More journals should offer manuscripts in both human-readable PDF and machine-readable XML like @MDPIOpenAccess does.

exceptional clinical trial title

Today we're publishing our index of unmet needs in human disease: 2443 indications scored and ranked on burden of disease, prevalence, pipeline activity, and treatment burden. We hope this will help drug developers identify overlooked medical problems convoke.bio/blog/introducing…

The "thrill of conjuring" - @OpenAI GPT-5.3-Codex-High just implemented bluetooth microphone recording functionality for an open source offline local voice notetaking & transcription app (NotelyVoice) on my phone, and I know absolutely nothing about android, kotlin, or bluetooth.

Because of all the refusals (via API and separately in chat) I get significantly lower performance on my biotech benchmark that covers antibody therapeutics, tox, clinical dev, cell surface receptors, food allergens, and biological sequences.

It's time for our 2026 Q1 update! Below is a list of all of the new platform features and improvements, including a free tier! If you haven't signed up or logged in recently, it's a great opportunity to start tracking trials that matter to you. 1/3

Had a chuckle at this one: "You are absolutely right—my apologies for the mix-up! 5V8S is indeed a bacterial protein from Thermotoga maritima and has nothing to do with..."

Nice introductory tutorial from @btnaughton on de novo antibody design. One comment on specificity: TCR (tissue cross reactivity) or even human proteome arrays are common here, rather than selecting individual proteins for off-target screening on a platform like adaptyv.

Many aspects of biotech involve handling long nucleotide and amino acid sequences, which, in my experience, LLMs do poorly. AutoBE as @SamchonGithub describes: "uses function calling. Instead of generating free-form text, the LLM fills in predefined structures — JSON Schema... Once the LLM fills the form, compilers validate and transform it" I think there is an analogy here for improving biological sequence handling performance by avoiding LLMs writing such sequences free-form, and additionally by having validators for sequences e.g. "B" isn't an amino acid.

Congrats to the Denali team

LAMBS (Local Analysis of mAb Sequences) now has multiple sequence alignment & analysis! Check it out: permissively licensed open source, dependency-free, and completely private!