@fabriziomaniglio.bsky.social
AI euphoria, meet compliance reality. Happy Monday.
October 5, 2026 at 6:30 AM
AI euphoria, meet compliance reality. Happy Monday.
Since 2024, life sciences got much better at experimenting with AI, and barely better at governing it.
Adoption sprinted. Governance stood still. Call it governance debt; the bill comes due at inspection.
Who owns paying yours down?
Adoption sprinted. Governance stood still. Call it governance debt; the bill comes due at inspection.
Who owns paying yours down?
September 30, 2026 at 12:03 PM
Since 2024, life sciences got much better at experimenting with AI, and barely better at governing it.
Adoption sprinted. Governance stood still. Call it governance debt; the bill comes due at inspection.
Who owns paying yours down?
Adoption sprinted. Governance stood still. Call it governance debt; the bill comes due at inspection.
Who owns paying yours down?
The augmentation-versus-replacement debate, in one Monday meme.
September 28, 2026 at 6:30 AM
The augmentation-versus-replacement debate, in one Monday meme.
AI can now generate more documentation than anyone will ever read. The reviewer, however, is one very tired cat. Happy Monday.
September 21, 2026 at 6:30 AM
AI can now generate more documentation than anyone will ever read. The reviewer, however, is one very tired cat. Happy Monday.
Everyone's running an AI pilot. Almost nobody has scaled one.
The problem isn't pilots. It's pilots with no exit: no real question, no KPIs, no path to production.
A pilot without a route to scale is just a nice way to feel busy.
The problem isn't pilots. It's pilots with no exit: no real question, no KPIs, no path to production.
A pilot without a route to scale is just a nice way to feel busy.
September 16, 2026 at 12:03 PM
Everyone's running an AI pilot. Almost nobody has scaled one.
The problem isn't pilots. It's pilots with no exit: no real question, no KPIs, no path to production.
A pilot without a route to scale is just a nice way to feel busy.
The problem isn't pilots. It's pilots with no exit: no real question, no KPIs, no path to production.
A pilot without a route to scale is just a nice way to feel busy.
AI hallucinations, illustrated. The citation looked real. The source was imaginary. Verify before you trust. Happy Monday.
September 14, 2026 at 6:30 AM
AI hallucinations, illustrated. The citation looked real. The source was imaginary. Verify before you trust. Happy Monday.
Oversight is only oversight if the human has time to actually look. Otherwise it is just... this. Happy Monday.
September 7, 2026 at 6:30 AM
Oversight is only oversight if the human has time to actually look. Otherwise it is just... this. Happy Monday.
A contradiction I can't stop thinking about: validation is the #1 area life sciences expects AI spend to grow (80%), yet half aren't even using GenAI there.
The blocker isn't the model. It's integration, regulatory clarity, and data governance.
The blocker isn't the model. It's integration, regulatory clarity, and data governance.
September 1, 2026 at 6:30 AM
A contradiction I can't stop thinking about: validation is the #1 area life sciences expects AI spend to grow (80%), yet half aren't even using GenAI there.
The blocker isn't the model. It's integration, regulatory clarity, and data governance.
The blocker isn't the model. It's integration, regulatory clarity, and data governance.
The eternal love triangle of every AI procurement cycle. Happy Monday.
August 31, 2026 at 6:30 AM
The eternal love triangle of every AI procurement cycle. Happy Monday.
In 2013, a machine guessed whether you were dreaming of a person or a place. Barely above a coin flip.
By 2023 it could rebuild the video you were watching from brain activity alone, about 85% accurate.
It reconstructs what you're seeing, not what you're imagining. That gap is the frontier.
By 2023 it could rebuild the video you were watching from brain activity alone, about 85% accurate.
It reconstructs what you're seeing, not what you're imagining. That gap is the frontier.
August 25, 2026 at 12:02 PM
In 2013, a machine guessed whether you were dreaming of a person or a place. Barely above a coin flip.
By 2023 it could rebuild the video you were watching from brain activity alone, about 85% accurate.
It reconstructs what you're seeing, not what you're imagining. That gap is the frontier.
By 2023 it could rebuild the video you were watching from brain activity alone, about 85% accurate.
It reconstructs what you're seeing, not what you're imagining. That gap is the frontier.
Decision integrity is really just the art of never having to be this guy in front of an inspector.
August 24, 2026 at 6:30 AM
Decision integrity is really just the art of never having to be this guy in front of an inspector.
DeepMind's Perch 2.0 was trained on 1.5M bird recordings, 14,597 species. Virtually zero underwater data.
Pointed at orca calls, it scored 0.945. The purpose-built whale classifier scored 0.821.
Transfer learning at its most vivid: the same bet pharma makes with LLMs on batch records.
Pointed at orca calls, it scored 0.945. The purpose-built whale classifier scored 0.821.
Transfer learning at its most vivid: the same bet pharma makes with LLMs on batch records.
August 19, 2026 at 12:03 PM
DeepMind's Perch 2.0 was trained on 1.5M bird recordings, 14,597 species. Virtually zero underwater data.
Pointed at orca calls, it scored 0.945. The purpose-built whale classifier scored 0.821.
Transfer learning at its most vivid: the same bet pharma makes with LLMs on batch records.
Pointed at orca calls, it scored 0.945. The purpose-built whale classifier scored 0.821.
Transfer learning at its most vivid: the same bet pharma makes with LLMs on batch records.
The industry just drew a line for AI, and it's strikingly consistent: adoption falls, readiness falls, concern rises the moment AI decides on its own.
Not a ban. A boundary. Exactly what Annex 22 says.
Where's your line?
Not a ban. A boundary. Exactly what Annex 22 says.
Where's your line?
August 18, 2026 at 6:30 AM
The industry just drew a line for AI, and it's strikingly consistent: adoption falls, readiness falls, concern rises the moment AI decides on its own.
Not a ban. A boundary. Exactly what Annex 22 says.
Where's your line?
Not a ban. A boundary. Exactly what Annex 22 says.
Where's your line?
QA has survived every "this will replace you" since the fax machine.
The coffee helps. ☕
The coffee helps. ☕
August 17, 2026 at 6:30 AM
QA has survived every "this will replace you" since the fax machine.
The coffee helps. ☕
The coffee helps. ☕
Your voice changes before your hands start shaking.
AI trained on vocal biomarkers can detect early Parkinson's from a 30-second voice recording. 91.11% accuracy, peer-reviewed (Scientific Reports).
The gap is validation at scale, not technology.
AI trained on vocal biomarkers can detect early Parkinson's from a 30-second voice recording. 91.11% accuracy, peer-reviewed (Scientific Reports).
The gap is validation at scale, not technology.
August 12, 2026 at 12:02 PM
Your voice changes before your hands start shaking.
AI trained on vocal biomarkers can detect early Parkinson's from a 30-second voice recording. 91.11% accuracy, peer-reviewed (Scientific Reports).
The gap is validation at scale, not technology.
AI trained on vocal biomarkers can detect early Parkinson's from a 30-second voice recording. 91.11% accuracy, peer-reviewed (Scientific Reports).
The gap is validation at scale, not technology.
AI can draft my validation strategy in minutes. Moving one image in Word still detonates the entire document.
Our priorities are clearly in order. Happy Monday.
Our priorities are clearly in order. Happy Monday.
August 10, 2026 at 6:30 AM
AI can draft my validation strategy in minutes. Moving one image in Word still detonates the entire document.
Our priorities are clearly in order. Happy Monday.
Our priorities are clearly in order. Happy Monday.
Why DRIP is the real barrier to AI in validation: about half of life sciences orgs rate their own data only "good," not "excellent." And it gets worse the more autonomy you ask for.
AI doesn't fix a data-integrity problem. It inherits it.
AI doesn't fix a data-integrity problem. It inherits it.
August 7, 2026 at 12:02 PM
Why DRIP is the real barrier to AI in validation: about half of life sciences orgs rate their own data only "good," not "excellent." And it gets worse the more autonomy you ask for.
AI doesn't fix a data-integrity problem. It inherits it.
AI doesn't fix a data-integrity problem. It inherits it.
Casey Harrell has ALS. He lost his voice. A brain implant gave it back.
UC Davis: 97.5% word accuracy, sustained over 8 months and 84 sessions. Synthetic voice modeled on his pre-ALS recordings. Published in NEJM.
FDA-regulated trials, not demos. A new medical device category is being born.
UC Davis: 97.5% word accuracy, sustained over 8 months and 84 sessions. Synthetic voice modeled on his pre-ALS recordings. Published in NEJM.
FDA-regulated trials, not demos. A new medical device category is being born.
August 5, 2026 at 12:02 PM
Casey Harrell has ALS. He lost his voice. A brain implant gave it back.
UC Davis: 97.5% word accuracy, sustained over 8 months and 84 sessions. Synthetic voice modeled on his pre-ALS recordings. Published in NEJM.
FDA-regulated trials, not demos. A new medical device category is being born.
UC Davis: 97.5% word accuracy, sustained over 8 months and 84 sessions. Synthetic voice modeled on his pre-ALS recordings. Published in NEJM.
FDA-regulated trials, not demos. A new medical device category is being born.
A term that deserves word of the year in life sciences: DRIP. Data Rich, Intelligence Poor.
We have more data than any industry, and less usable intelligence from it than almost anyone.
Does it describe your org? Be honest.
We have more data than any industry, and less usable intelligence from it than almost anyone.
Does it describe your org? Be honest.
August 4, 2026 at 6:30 AM
A term that deserves word of the year in life sciences: DRIP. Data Rich, Intelligence Poor.
We have more data than any industry, and less usable intelligence from it than almost anyone.
Does it describe your org? Be honest.
We have more data than any industry, and less usable intelligence from it than almost anyone.
Does it describe your org? Be honest.
Nothing says "data-driven" like a beautiful dashboard balanced on a data swamp nobody has touched since 2015.
Happy Monday.
Happy Monday.
August 3, 2026 at 6:30 AM
Nothing says "data-driven" like a beautiful dashboard balanced on a data swamp nobody has touched since 2015.
Happy Monday.
Happy Monday.
60-73% of enterprise data goes unused for analytics. In pharma QC labs, probably worse.
Every chromatography run generates a full data stream. You use the pass/fail result and archive the rest.
Before buying another AI platform, ask: what would you find if you looked at what you already have?
Every chromatography run generates a full data stream. You use the pass/fail result and archive the rest.
Before buying another AI platform, ask: what would you find if you looked at what you already have?
July 30, 2026 at 12:02 PM
60-73% of enterprise data goes unused for analytics. In pharma QC labs, probably worse.
Every chromatography run generates a full data stream. You use the pass/fail result and archive the rest.
Before buying another AI platform, ask: what would you find if you looked at what you already have?
Every chromatography run generates a full data stream. You use the pass/fail result and archive the rest.
Before buying another AI platform, ask: what would you find if you looked at what you already have?
Everyone suddenly has a new skill: sounding fluent with AI.
In regulated work that's quietly dangerous. A fluent summary of a deviation is not a correct root cause. Sounding right and being right are not the same thing.
Stay fluent. Just don't confuse it with knowing.
In regulated work that's quietly dangerous. A fluent summary of a deviation is not a correct root cause. Sounding right and being right are not the same thing.
Stay fluent. Just don't confuse it with knowing.
July 29, 2026 at 12:02 PM
Everyone suddenly has a new skill: sounding fluent with AI.
In regulated work that's quietly dangerous. A fluent summary of a deviation is not a correct root cause. Sounding right and being right are not the same thing.
Stay fluent. Just don't confuse it with knowing.
In regulated work that's quietly dangerous. A fluent summary of a deviation is not a correct root cause. Sounding right and being right are not the same thing.
Stay fluent. Just don't confuse it with knowing.
A gut bacterium from a Japanese tree frog wiped out colorectal tumors in every mouse tested with one injection. It colonizes the tumor's low-oxygen core and pulls the immune system in. Cured mice later rejected new tumors, like a vaccine.
Early days and mouse-only, but beautiful biology.
Early days and mouse-only, but beautiful biology.
July 28, 2026 at 11:02 PM
A gut bacterium from a Japanese tree frog wiped out colorectal tumors in every mouse tested with one injection. It colonizes the tumor's low-oxygen core and pulls the immune system in. Cured mice later rejected new tumors, like a vaccine.
Early days and mouse-only, but beautiful biology.
Early days and mouse-only, but beautiful biology.
In 2023 a UC Davis/SETI team played a humpback contact call. A whale named Twain answered 36 times in 20 minutes, matching tempo.
AI found structure in what sounded like repetition. Same as an LLM finding deviation trends across a decade of batch records. Verifying what it means is still our job.
AI found structure in what sounded like repetition. Same as an LLM finding deviation trends across a decade of batch records. Verifying what it means is still our job.
July 28, 2026 at 12:02 PM
In 2023 a UC Davis/SETI team played a humpback contact call. A whale named Twain answered 36 times in 20 minutes, matching tempo.
AI found structure in what sounded like repetition. Same as an LLM finding deviation trends across a decade of batch records. Verifying what it means is still our job.
AI found structure in what sounded like repetition. Same as an LLM finding deviation trends across a decade of batch records. Verifying what it means is still our job.
Most "failed" AI pilots in life sciences didn't fail because the AI was bad. They failed because of where it was placed.
Bolt a capable model onto a broken workflow and you get faster mess.
The technology is fine. The wiring is wrong. Change the track, not just the engine.
Bolt a capable model onto a broken workflow and you get faster mess.
The technology is fine. The wiring is wrong. Change the track, not just the engine.
July 10, 2026 at 12:02 PM
Most "failed" AI pilots in life sciences didn't fail because the AI was bad. They failed because of where it was placed.
Bolt a capable model onto a broken workflow and you get faster mess.
The technology is fine. The wiring is wrong. Change the track, not just the engine.
Bolt a capable model onto a broken workflow and you get faster mess.
The technology is fine. The wiring is wrong. Change the track, not just the engine.