Shaping the future of digital health policy ๐๐๐
London
Joined June 2023
- Tweets1.9K
- Following258
- Followers25
- Likes6.6K
Dolores HealthTech retweeted
Chris Camillo, the social arbitrage investor profiled in Jack Schwager's Unknown Market Wizards, describes his method as spotting change in the world a little quicker than everyone else.
"The larger the change, the more opportunity," he says. The faster the change, the larger the opportunity.
No fundamental analysis. No technical analysis. "You're not even looking at stock price."
Dolores HealthTech retweeted
Anthropic's IPO will be ushered in with a whole bunch of model releases to celebrate AI
- Carbon will become generally available
- Anthropic will drop Fable and Opus variants
- OpenAI will match with Astra
AI will continue to accelerate, and I am sorry, there is nothing anyone can do about it ๐
Dolores HealthTech retweeted
AAV gene therapy insertional cancer cases (late 2026 update):
โก๏ธ Zolgensma (AAV9 for SMA): High-dose intravenous (1.1 ร 10ยนโด vg/kg) given systemically to reach the CNS. Among >6,000 treated children, ~9 tumors have been reported (~0.15%). At least 2โ3 of these showed clear vector integration that activated the GLI1 gene via promoter insertion. Tumors appeared in the CNS, spinal/paraspinal region, peripheral nerves, and kidney, typically 1.5โ5 years later.
โก๏ธ FBX-101 (AAVrh10 for Krabbe): Intravenous dose of 3 ร 10ยนยณ vg/kg given after HSCT to treat both the central and peripheral nervous systems. One case developed a paraspinal tumor. A truncated piece of the vector (strong CAG promoter plus a splice donor, missing most of the therapeutic gene) integrated into the GLI1 gene and drove a cancerous fusion transcript.
โก๏ธ RGX-111 (AAV9 for MPS I): Delivered directly into the CSF around the brain (intracisternal, ~1โ5 ร 10ยนโฐ genome copies per gram of brain mass). One patient developed a neuroepithelial brain tumor 4 years later. Vector sequences containing a strong CBA promoter fused to the PLAG1 gene, producing a chimeric transcript that overexpressed this oncogene.
โ ๏ธ GLI1 and PLAG1 are both developmental transcription factors. In young children their target progenitor cells are still active, so a strong viral promoter landing nearby can quickly drive a clonal tumor.
๐ The same insertions are less likely to cause rapid tumors once these developmental programs are silenced in adults and adolescents.
๐จ The early pediatric cases already prove that AAV promoter insertion can drive cancer. Longer-term risk from other gene hits is therefore a real concern that is likely to emerge as treated patients age.
โ๏ธ Despite all these, AAV vectors remain an acceptable benefit-risk option for many severe genetic diseases, since the alternative is far worse.
Dolores HealthTech retweeted
AgenticAI is all the rage, but orchestration, credentials, governance, and the cessation of AI agents' activities are real challenges and massive cybersecurity risks.
A look at what @IBM is doing in this space.
๐ newsroom.ibm.com/2026-10-27-โฆ
#ResponsibleAI
Dolores HealthTech retweeted
People with mental disorders can face disparities in #diabetes care. How can health services respond? ๐น ๐
This commentary by Olga Kozlowska and Louise Nicholson considers the role of redesigned care delivery and better health information systems๐ bit.ly/3WJ1Pr1
Dolores HealthTech retweeted
How would you feel if you found out that your physician uploaded your medical records to a large language model?
A STAT investigation examines how hospitals are responding to physicians' growing use of free clinical AI assistants.
One example involves Monument Health, a healthcare system operating in the US. According to the hospital's CIO, a physician had stored more than 100 patient records in OpenEvidence and was using the system to assist with medical decisions.
The hospital decided to block access to the platform.
Clinicians increasingly use AI systems that may operate outside their hospitals' approved technology infrastructure. And we have to talk about it.
Vendors including OpenEvidence claim HIPAA-compliant arrangements are available, but such claims do not automatically establish that every use complies with hospital policies or applicable privacy requirements.
Source: statnews.com/2026/11/05/doctโฆ
Dolores HealthTech retweeted
ADIB Adds 24/7 Transfers to Egyptian Bank Accounts via Visa Direct fintechnews.ae/34787/paymentโฆ #middleeast #fintech
Dolores HealthTech retweeted
A model being great at US English says little about how it ranks in other languages like Brazilian Portuguese.
These benchmark efforts matter a lot for voice.
The per-use-case benchmark is also ideal for properly measuring voice agents' effectiveness and robustness in certain languages.
Drop 3/15: Introducing the Speech Recognition Benchmark on @voicearena_ai.
One benchmark for every use case of speech recognition, in every language. Tested on real conversations.
Dataset 1 is live: the Monsoon Conversational ASR Test Set, two-person conversations in 5 languages.
Leaderboard and public test sets at Voice Arena.
Dolores HealthTech retweeted
Brilliant paper on optimizing agent environments.
This is an extremely important skill if you work with agents today.
You can make a lot of improvements by optimizing the environments agents operate in.
Many agents work inside a shared workspace, such as a company's Slack or a set of documents and tickets. EnvOpt changes what that workspace holds and how it is presented to the agent.
A coding agent runs the target agent on training tasks without graders or reference answers, edits the environment, and stops based on a validation set.
Haiku 4.5's pass rate also rises by 14.34%. The same environment transfers to Opus 5 with 68.1% lower cost. On top of Meta-Harness's optimized harness, it lowers cost by another 83.4% and adds 11.17% accuracy.
On AutomationBench, it cut Claude Haiku 4.5's cost by 72.7%.
arxiv.org/abs/2611.07372
Chat with Paper: academy.dair.ai/papers/envopโฆ
Dolores HealthTech retweeted
Weight memory vs. retrieval in agents. Brilliant work from Amazon.
This work shows weight memory can outperform text memory on recall of user facts.
But the best part is that you can leverage both.
Using the same model and user data, this work shows that plain fine-tuning stores user facts in weights at 36.9% accuracy, while on-policy self-distillation reaches 59.9%.
That second number also beats retrieval.
On eight HaluMem-Long users, the distilled model answers with no retrieved context and beats MemOS-style retrieval (47.8%), extracted context (51.9%), and the raw dialogue (51.3%). The gain holds for every user and at 4B, 9B, and 27B.
How it works: a teacher, which is the same model reading the user's memory, trains a student that has to answer without it.
Weight memory has clear limits. It forgets older facts and cannot tell when something was never recorded. A verify-then-regenerate router checks each answer against the text store and adds 6.2 points.
The recommendation is to keep a text store for recent and missing facts, and move stable user knowledge into weights.
arxiv.org/abs/2611.05636
Chat with Paper: academy.dair.ai/papers/routeโฆ
Dolores HealthTech retweeted
high CPU usage in a .NET application doesn't always mean the business logic is doing too much work.
excessive allocations can increase garbage collection overhead and consume CPU.
sometimes reducing allocations improves performance more than optimizing algorithms.
Dolores HealthTech retweeted
You can't call this AI slop anymore.
One self-taught working mom made this 9-minute AI fantasy film.
An she won 4 awards at the New York International Film Awards, including Best AI Film and Best Screenplay.
This is what real AI filmmaking looks like ๐
Dolores HealthTech retweeted
Today we release Ripple-65B Preview, our 65B dense LLM we pretrained from scratch.
Ripple sets a new state-of-the-art for U.S. open-weight models in agentic coding tasks, surpassing significantly larger models including Muse Spark Open-816B, Beam-501B, Inkling-975B, and Nemotron 3 Ultra-550B.
Learn more: ricursive.com/ripple
Dolores HealthTech retweeted
Impressive paper from CMU.
Soon, we will have coding agents that can tune the entire stack, from prompts to token serving.
This work experiments with coding agents to tune an LLM serving stack.
They get Claude Sonnet 5.5 to build a serving system for Gemma E4B that delivers 4.87x the value per GPU dollar of the vLLM reference.
ServerBench asks coding agents to design a complete, specialized serving system and decide how many concurrent sessions to run. The score is latency-qualified token value per dollar of GPU cost, measured by replaying real coding sessions with their dependencies intact. An independent check verifies the system works.
Across 11 coding-agent models and three deployments, 22 of 33 runs produced a verified system and 13 beat the strongest baseline.
The best designs raised capacity and chose the load to match it. Failed runs often shipped optimizations whose new code paths were never tested.
arxiv.org/abs/2611.06841
Chat with Paper: academy.dair.ai/papers/serveโฆ
Dolores HealthTech retweeted
Two surgeons take on healthcareโs version of Minority Report: can AI predict a patientโs future?
Live from the Second Opinion Summit, Dr. Jay Shah (CMO of @concurrencehq, Stanford cancer surgeon) and @Dr_PJay (Senior Medical Director at Optum Insight, orthopedic surgeon) argue that the data that matters most lives in the 99.9% of a patientโs life that happens between appointments. We also get into what a digital twin really is, why 20โ25% of knee replacement patients are unhappy, what ambient scribes leave out, and more.
Timestamps below.
(0:00) Intro
(1:10) Healthcareโs Minority Report
(2:21) Meet Jay and Prakash
(5:25) Jayโs brush with septic shock
(6:56) From surgeon to systems thinker
(10:20) Donโt doctors already predict the future?
(14:07) Why knee replacement patients are unhappy
(17:51) What a digital twin really is
(20:57) The other 99.9% of a patientโs life
(24:15) What ambient scribes leave out
(27:20) The medical record isnโt the source of truth
(28:03) Privacy and lessons from social media
(30:30) Data, prediction, decision intelligence
(35:00) Sharing data and the learning loop
(38:00) Learning health systems and the M&M
(41:03) Will AI replace doctors?
(43:30) The 82-year-old patient
(46:58) Whatโs next for AI prediction
(48:20) Wrap
Dolores HealthTech retweeted
"That gap between 'it works on my laptop' and 'it works on Monday morning with real data' has picked up a name: harness engineering."
Gursimar Singh presents a clear and accessible introduction to the components of production-grade agentic AI. towardsdatascience.com/your-โฆ
Dolores HealthTech retweeted
Oaktree Capital co-chair Howard Marks calls AI the โleast knowable, least predictableโ development of his investing career.
He says it's nearly impossible to forecast which companies will win or what their AI products will be worth in five years.
(Source: @CNBC)
Dolores HealthTech retweeted
The irony about patient consent and human oversight in fast-moving, acute care, life-or-death environments is worth greater discussion.
teapaw.org/AI_4_Healthcare/statusโฆ
#ResponsibleAI
Oh, the irony ... how little U.S. patients know. ๐ค
60% ๐ผ๐ณ ๐๐บ๐ฒ๐ฟ๐ถ๐ฐ๐ฎ๐ป๐ ๐ฎ๐ฟ๐ฒ ๐๐ป๐ฒ๐ฎ๐๐ ๐๐ถ๐๐ต ๐๐ ๐ฑ๐ถ๐ฎ๐ด๐ป๐ผ๐๐ฒ๐.
๐ ๐ฎ๐ป๐ ๐ผ๐ณ ๐๐ต๐ฒ๐บ ๐ต๐ฎ๐๐ฒ ๐ฎ๐น๐ฟ๐ฒ๐ฎ๐ฑ๐ ๐ฏ๐ฒ๐ฒ๐ป ๐ฐ๐ฎ๐ฟ๐ฒ๐ฑ ๐ณ๐ผ๐ฟ ๐ฏ๐ ๐ถ๐.
New @pewresearch (released October 29th) finds that 60% of Americans are uncomfortable with their doctor using AI for diagnoses and treatments which is twice the 30% figure of those who are comfortable.
So let's follow a typical metropolitan hospital stay from start to finish, shall we ...
๐ค For scheduled procedures, intake via chatbot.
๐ In the ER, AI-triage risk scores flag who can't wait.
๐ฉบ Clinicians using AI-enable stethoscopes.
๐ฉธ Your bloodwork assays are read by AI to check for biomarkers and look for values outside healthy ranges.
๐ฉป Algorithms look at X-rays, MRIs, and other imaging scans to push a suspected stroke or bleed to the top of the radiologist's queue.
โค๏ธ ECG software reads the trace before the cardiologist does.
๐ Pharmacy software and tools run potential drug interactions and dosing checks on every Rx.
โ ๏ธ In the ICU and wards early warning models watch vitals for sepsis and deterioration.
๐ฉโโ๏ธ๐งโโ๏ธ Your healthcare team using AI-enabled clinical decision support tools on their phones to look up the latest research, ask questions about your case, and seek advice on your care plan.
๐ In discharge, readmission risk models shape follow-up and your at-home care notes are courtesy of a Generative AI model.
Most patients never know any of this happens, and a clinician stays in the loop at every step. So the question isn't AI vs human. It's how we make AI that is already critical to outcomes more visible, validated, and accountable.
This is why patient consent to AI in our care is becoming a MOOT, if not, NON-NEGOTIABLE, point.
The use of and expansion AI in your personal care is literally changing between each visit to your doctor, clinic, or hospital vists ... week by week!
๐pewresearch.org/science/2026โฆ
#HealthcareAI #PatientSafety
Dolores HealthTech retweeted
Americans Open to #AI in #Healthcare but Want Human Oversight
pewresearch.org/science/2026โฆ
#MedX #MedTwitter #MedEd #AIinHealthcare #HealthAI #HealthTech
Dolores HealthTech retweeted
CrowdStrike ($CRWD) is preparing for a new cybersecurity problem as companies deploy more autonomous AI agents.
The concern is that compromised or poorly controlled agents could operate inside corporate systems with access to sensitive data and tools.
benzinga.com/trading-ideas/lโฆ
