https://xitter.cf/t.co/E3iA6NjTGg PhD student in ML, computer vision & biology at Stanford 🇦🇺

Stanford, Ca
Joined May 2021
Check out PaperSearchQA, which I'll present at EACL in Morocco this March! We built an RL training environment for teaching LLMs to search and reason over scientific papers. 60k question-answer pairs + 16M papers to search over + benchmarks. RL training improves the model.
6
14
1
37
8,590
James Burgess retweeted
X-rays are everywhere in medicine, but extracting general, quantitative anatomical information from them remains remarkably difficult. Announcing FleXray: an open-source, open-weight model for zero-shot anatomical segmentation across the human body—from different regions, viewing angles, and acquisition settings. FleXray learns from large-scale synthetic supervision generated from 3D anatomy, then transfers directly to real clinical X-rays. Try it now on your own X-rays in our in-browser demo! 🧵 1/N — project page, demo, weights, code & paper below ↓
18
63
17
424
69,739
James Burgess retweeted
Yes those Silicon Valley bros will end up imposing “smart baby” selection. Sounds awful/immoral, until enough ambitious parents participate, and drag everyone into the ultimate arms race. Same escalation mechanism as elite schools, tutors, admissions consultants etc, except moved 9 months earlier. Most parents probably don't particularly want the optimisation arms race. But once enough ambitious parents participate, opting out starts feeling less like moral restraint and more like voluntarily disadvantaging your kid. The usual status mechanics kick in: Rich weirdos test something → successful/high-status people normalise it → upper-middle class copies it → opting out becomes the weird choice. Health will probably be how the whole thing becomes socially acceptable too. Nobody wants to say “we paid to make our baby smarter”. But screening embryos for lower Alzheimer’s, heart disease or diabetes risk sounds pretty reasonable. Then once you have 10 embryos ranked on a screen, I’m not sure humans will be very good at deliberately ignoring the height / intelligence / longevity columns next to it. And tbh rich parents already spend fortunes buying better probabilities for their children after birth (postcode, private school, tutors, sports coaching, internships, connections etc.) and everyone finds that normal. But somehow buying better odds before birth suddenly feels dystopian. The optimiser class was probably always going to end up here anyway, having spent the last 30 years moving the optimisation of childhood earlier and earlier: university consultants → tutors → elite nursery → prenatal testing → egg freezing/IVF. Embryo selection is just the point where competitive parenting finally reaches before the child exists.
35
52
16
630
41,348
RL-trained agents outperform the retrieval baselines (Qwen2.5-7B: 51.0% vs RAG 36.5%), and generalize to BioASQ, a human-created benchmark (+15.1 pts over RAG). We also see some interesting behaviours in the agent transcript (see the project page).
1
2
153
Thanks to my advisor @yeung_levy and collaborators, Jan, Duo, @Zhang_Yu_hui ,@Prof_Lundberg, @Ale9806_, @minwsun. Work done at @StanfordAILab All links at the project page: jmhb0.github.io/PaperSearchQ… And HuggingFace: huggingface.co/papers/2601.1…
1
2
163
RL-trained agents outperform the retrieval baselines (Qwen2.5-7B: 51.0% vs RAG 36.5%), and generalize to BioASQ, a human-created benchmark (+15.1 pts over RAG). We also see some interesting behaviours in the agent transcript (see the project page).
1
79
Thanks to my advisor @yeung_levy and collaborators, Jan, Duo, @Prof_Lundberg, @Ale9806_, @minwsun. Work done at @StanfordAILab All links at the project page: jmhb0.github.io/PaperSearchQ… And HuggingFace: huggingface.co/papers/2601.1…
1
89
James Burgess retweeted
Midway through grad school, I was at the Monterey Bay Aquarium dazzled, watching cuttlefish change the colours and textures on their skin. This morning, our paper, "Soft photonic skins with dynamic texture and colour control" (rdcu.be/eX1Sw) was published in Nature.
4
3
7
275
James Burgess retweeted
🛠️ How should an AI system develop new reasoning tools? The same way expert programmers do—by learning from experience, not guessing upfront. We introduce 𝐓𝐫𝐚𝐧𝐬𝐝𝐮𝐜𝐭𝐢𝐯𝐞 𝐕𝐢𝐬𝐮𝐚𝐥 𝐏𝐫𝐨𝐠𝐫𝐚𝐦𝐦𝐢𝐧𝐠 (𝐓𝐕𝐏), a framework that builds reusable tools from its own problem-solving experience. 🧵👇
2
6
17
1,060
James Burgess retweeted
We’re thrilled to announce that Wispr Flow raised another $25M round after 10x'ing our revenue in just 5 months. Our Series A2 was led by @hanstung at @notablecap (who was an early investor in five companies that made it to $100B valuation like Slack, Tiktok, and Airbnb). we also brought on @StevenBartlett as an investor and partner. But here's what matters more than the money: We cracked voice input. Not transcription - actual understanding. Our users hit "send" in under 0.5 seconds without checking. They trust it blindly. That's never existed before. In a recent benchmark, Wispr came out as 3-4x more accurate than OpenAI, ElevenLabs, and Siri. Voice input was step one. Now, we’re building the assistant that actually gets things done. The keyboard had a good 150-year run. Time to build what comes next. PS: like, retweet, and bookmark to get wispr flow for free for 3 months ❤️ — Written with @WisprFlow
73
110
21
458
99,509
What does it really mean for two videos to be similar? At #NeurIPS2025, we’ll present ConViS-Bench: Estimating Video Similarity Through Semantic Concepts. Stop by our poster for a chat! 📍Exhibit Hall C,D,E — Poster N.4618 🕒 Thu, Dec 4, 2025 • 11:00 AM – 2:00 PM PST
1
6
2
19
3,655
I'll be at NeurIps from Wed-Fri, and keen to chat with people about VLMs, agents, and applications to science.
1
1
10
488
James Burgess retweeted
Thinking about using small multimodal models? Want a clearer understanding of what breaks when downscaling model size, and why? ✨Introducing our new work on Downscaling Intelligence: Exploring Perception and Reasoning Bottlenecks in Small Multimodal Models 🧵👇
1
8
2
39
8,402
“Make it red.” “No! More red!” “Ughh… slightly less red.” “Perfect!” ♥️ 🎚️Kontinuous Kontext adds slider-based control over edit strength to instruction-based image editing, enabling smooth, continuous transformations!
16
36
3
161
29,193