PHDone @ UToronto | AI for biology & drug discovery
Toronto, Ontario
Joined April 2018
- Tweets93
- Following764
- Followers157
- Likes1.3K
Pinned Tweet
Excited to see X-Scientist launch! Proud to be one of the core builders of its harness alongside an incredible team @Xaira_Thera.
Here’s what building it taught me about domain-specific agents 👇
1, The model is general. The harness is where domain knowledge lives. Frontier models already “know” a lot of biology. What they lack is the scaffolding to think like a drug hunter: which evidence counts, what order to reason in, when to change course. Our key design choice was to separate scientific thinking from execution. Once coordination and verification are their own layer, they become something you can optimize, and that’s where domain knowledge gets encoded.
2, Tools > prompts. A drug discovery agent needs tools built for its domain. We built ours around our own models like X-Cell and X-Design, alongside bioinformatics and scientific analysis tools. Each tool has to return outputs the model can reason over, not raw dumps.
3, Your verifier is your moat. Coding has unit tests. Biology has noisy data and slow wet-lab loops. Deciding what “good” looks like at each step, and checking it automatically, is what keeps an agent grounded instead of drifting into confident nonsense. It’s also what lets scientists trust its output and step in where it matters.
4, The thinking layer learns. Because scientific reasoning is separated from execution, it can improve on its own. Every run leaves a trace of which strategies produced real evidence and which burned effort. We feed that back into how the agent plans, allocates and verifies. Each run makes the next one a better scientist.
General-purpose agents will keep getting better. But in high-stakes science, the last mile belongs to whoever builds the best domain harness. 🚀
Meet X-Scientist: our step toward a self-improving virtual drug hunter!
Today, we’re introducing Xaira’s agentic AI platform for drug discovery:
🧠 Beats Codex using the same Astra model by 7.4 percentage points on DrugDiscoveryBench.
🧬 Nominates targets ~10× faster than human curation, with a comparable wet-lab hit rate.
🔁 Conducts autonomous research, outperforming AutoResearch and AutoScientists in the PerturbQA benchmark.
Here’s what we built, what we tested, and where we’re going. 🧵
Rex Ma retweeted
The biggest bottleneck in AI for biology is still the wet lab.
AI can generate more ideas than our labs can test. Closing that gap is how we turn better models into faster discovery.
Today, we’re introducing X-Automate at Xaira: bringing agentic AI and physical robotics into the lab. 🧵
X-Automate + X-Scientist for closing lap in the loop 🧪
Meet X-Automate: our step toward a self-driving lab!
We’re connecting @Xaira_Thera's agentic AI to the physical lab:
🦾 X-Bot operates instruments built for people.
🧠 X-Scientist coordinates sample preparation, transport, and analysis.
⚙️ 80% more plates analyzed over seven days vs. a modeled weekday 9-to-5 loading schedule.
Here’s what we built, tested, and learned. 🧵
Full technical blog: xaira.com/news/x-scientist-r…
Rex Ma retweeted
We’re betting our $5.5M seed round that the next leap in AI coding won’t come from OpenAI or Anthropic.
It’ll come from a new feedback loop that learns from you.
Introducing Beacon: open-source, self-improving memory for coding agents.
Your sessions capture how you debug and solve problems. Today, that context gets lost and you have to re-explain things every time you switch tools.
Beacon continuously distills your sessions into reusable skills across Claude Code, Cursor, Codex, and 20+ harnesses, so your next agent can learn from the last.
Beacon is free and OSS 🌟 github.com/Asymptote-Labs/ag…
Why should the frontier labs be the only ones winning with your sessions?
You already did the work. Now let your coding agents learn from you.
Rex Ma retweeted
Cost and speed are the small prizes here , yet sovereignty is the big one. A pharma renting frontier intelligence is strategically exposed: access, pricing, and roadmap decided elsewhere, with its most sensitive data crossing a boundary it doesn't control. Meanwhile every major pharma already owns the one asset frontier labs can't replicate : decades of proprietary screens, assay readouts, and failed trials. Fine-tune open weights on that data and you own the whole stack: weights, inference, IP. That's AI sovereignty at company scale. The first pharma to internalize this compresses discovery timelines by years. The rest end up renting intelligence about their own molecules.
We were able to fine tune an open weight model (Qwen3.5-9B) on proprietary internal data at @coinbase, and it out performed frontier models, at much lower cost and speed, for a specialized task (fraud prevention in this case).
I think we'll see more of this pattern inside companies. The frontier models are amazing generalists. But increasingly you'll see specialist models out perform on more narrow tasks, either working on teams, or by themselves.
We're making it easy to fine tune models internally for our employees. We even have a skill for it in our internal harness that makes it easy to spin these up and start training them:
"/fine-tune-and-serve-llm I want to fine-tune Qwen for payments risk review. help me get started"
I imagine we'll have some backend services that could be replaced entirely by this new architecture.
Rex Ma retweeted
We’ve been quiet about protein design at @Xaira_Thera . Here’s what we’ve been building:
Introducing X-Design, with results from two active drug programs that show what we think this field needs to deliver: speed to a useful lead, and access to targets that have resisted discovery.
→ 7 weeks from DNA synthesis to a preclinical lead on an oncology target, with the epitope and reagents already prepared.
→ A functional antibody antagonist against a challenging GPCR where conventional library screening AND llama immunization had failed.
The bar we set for our model goes well beyond binding.
The antibodies need to be human-like, functional, cross-reactive across species, and developable. Properties that have to come together in the same molecule for a program to move forward.
That’s what makes these results exciting to me: designs tested in the lab, meeting the requirements of real programs in our pipeline.
My bet is that the biggest impact of AI protein design will be expanding which diseases and targets we can pursue.
We want to choose targets because the biology is compelling and to build the ability to make the right molecules.
This is our first look at X-Design. Much more to come.
See our blog: xaira.com/news/de-novo-desig…
Rex Ma retweeted
Nearly half a year of silence. We spent it studying one problem: how far RL can scale.
MiMo-V2.6 is in the middle of its RL run right now. Three things we scaled: compute (~2B tokens per step, 1568 prompts × 16 rollouts, fully async), environments and harnesses (multi-task agentic RL, mixed across multiple harnesses in one run), and grader compute (agentic in-group credit assignment, with test-case and rubric-based rewards). We'll open-source the details piece by piece over the coming weeks.
Streaming the run: mimo.xiaomi.com/rl/
Rex Ma retweeted
Harnesses often get dismissed as just scaffolding, just prompt engineering, and not real research. But that couldn't be farther from the truth. The same model weights that score 30% on ARC-AGI score 95% with a better harness.
So we gathered a group of researchers and founders working at the frontier to do a deep dive into the state of harnesses.
We cover how we got to this point, the case for making your harness as expressive as possible, and what YC learned building an agent for every employee in the company.
00:00 - @FrancoisChauba1: Why harnesses matter
04:27 - Building an auto-researcher by accident
07:13 - A five minute history of harnesses
13:56 - Self-improving harnesses
18:35 - @sethkarten: Prime Agent, a self-improving RLM harness
21:50 - Context as an L1, L2, L3 cache
24:51 - From Turing machine to von Neumann computer
28:33 - Messaging between agents
30:04 - ARC-AGI results
33:09 - Emulator Bench and GPU kernels
37:30 - @JonSaadFalcon: OpenJarvis, personal AI on personal devices
38:26 - How far behind are local models
39:21 - The five primitives of a personal AI stack
42:47 - Letting cloud models optimize your local stack
43:53 - 800x cheaper than the cloud
45:58 - @josh__france and @jbellregan: QM, YC's agent harness for work
47:29 - A history of YC's internal agents
49:24 - OpenClaw and a fleet of 50 agents
51:04 - Pulling the brain out of the sandbox
54:43 - Letting the agent choose its own sandbox and model
57:16 - The grind tool: budgets on goals
58:50 - Agents don't understand social context
Rex Ma retweeted
GPT-6 Astra, trained on ~100K+ NVIDIA Grace Blackwell NVLink72. From ChatGPT to o1 to Astra in 4 years.
AGI has arrived. Congratulations @OpenAI team.
400K GPUs coming online next.
Rex Ma retweeted
Big news: Kimi-K3 by @Kimi_Moonshot is now #1 in the Frontend Code Arena with 1679 pts, surpassing Claude Fable 5.
This is a 17-place jump from Kimi-k2.6 (#18 -> #1).
In Frontend, Kimi-K3 ranked #1 in 6 of 7 domains: Brand & Marketing, Reference-Based Design, Data & Analytics, Consumer Product, Simulations, and Content Creation Tools, landing #2 only in Gaming behind Fable 5.
The full model weights will be released by July 27.
Congrats to the @Kimi_Moonshot team on this major milestone!
More AI investment should go to Canada!
We’re committing $10 million CAD and partnering with leading AI institutions in Canada to help fund new AI research.
anthropic.com/news/canadian-…
Rex Ma retweeted
With our internal coding benchmark, we're able to confidently introduce open-weight models into our AI code reviewer w/o degrading code quality. Have the frontier model (Fable) to the hardest work, delegate lower-level work to Kimi K2.6
Better quality, cheaper cost.
Because of DashBench, we’re able to quickly discern which model combinations yield the best results and at the best cost.
For example, with DashBench we’ve seen Kimi K2.6 + Fable 5 vastly outperform our current Sonnet 4.6 + Opus 4.8 harness at a cheaper cost. Having our own benchmark allows us to build confidence in leveraging the frontier intelligence and open-weight models without compromising on enterprise outcomes.
ALT Weighted Precision/Recall Tradeoff chart on DashBench
ALT Chart showing Kimi K2.6 + Fable 5 vs. Sonnet 4.6 + Opus 4.8 code reviewer harnesses
Rex Ma retweeted
new post on harness engineering for AI self-improvement: lilianweng.github.io/posts/2…
It is hard to forecast how much the future of RSI will rely on harnesses. Likely harness engineering will evolve in the direction of self-improvement and enable auto-research, and, in turn, smarter models keeps harnesses simple.
Even when many harness improvement get eventually internalized into core model, the need to specify goals and context will not disappear.
Rex Ma retweeted
🚀 Excited to share Xaira Therapeutics' first scientific blog post!
AI has transformed our ability to generate antibody binders. But generating more molecules isn't the end goal.
The real challenge is identifying the molecules with the greatest potential to become medicines.
In this post, Hetu (CTO) and Debbie Law (CSO) share how we're thinking about progressable binders: shifting the focus from simply finding binders to designing and selecting candidates that can successfully move through the drug discovery process.
I'm especially proud that this is Xaira's first scientific blog (with more to come!!). It's a great example of how we're combining AI, experimental science, and decades of drug development experience to tackle problems that matter.
Congratulations to Hetu, Debbie, and everyone who contributed!
Read more: xaira.com/news/progressable-…
Rex Ma retweeted
As engineering, product, design, DS, etc. melt into a new kind of role, I was reflecting on what roles might look like in the future. For example, when I look at the Claude Code team I see what I think is five archetypes:
1. Prototyper: comes up with brand new ideas; churns out many ideas, most of which don't ship
2. Builder: quickly turns a prototype/idea into production-grade product/infra
3. Sweeper: cleans up the UI, simplifies the code and system, unships, optimizes performance
4. Grower: takes a product that has been built and iterates on it to improve Product-Market Fit
5. Maintainer: owns a mature system to make it secure, reliable, fast, and efficient as it scales
Many people span across 2 roles, and sometimes 3 roles. I also notice that these roles are not really tied to job function -- eg. across Anthropic, some designers match category 1, some 2, some 3; same for engineers, PM, DS.
A healthy team needs a mix of these, depending on the product:
- A product that is new and pre-PMF needs people that are strong at 1+2+3
- A product that is growing and has found PMF needs 2+3+4 and some 5
- A product that has strong PMF needs 3+4+5 and some 2
Maybe product roles of the future will look more like this, and less like the domain-specific roles of today?
Rex Ma retweeted
📣📣 Meet Qwen-AgentWorld — a native language world model that simulates 7 agent environments (MCP, Search, Terminal, SWE, Web, OS, Android) within a single model. Environment modeling is the training objective from day one, not a post-hoc adaptation.
🤔 LLMs are trained to be better agents — better at acting in environments. But nobody has trained them to model the environments themselves.
🗺️ Our roadmap: investigate how language world modeling can push the boundaries of general agent capabilities, along two routes:
1️⃣ Build a foundation model for environment simulation — outperforming Claude Opus 4.8 and GPT-5.4 on AgentWorldBench
2️⃣ Investigate how world modeling enhances agent training:
🔬 Controllable Sim RL (agentic RL with LWM as environments) surpasses training in real environments
🧠 Learning to predict environments (LWM warm-up) makes agents stronger — remarkably, even without any agent-specific training, this predictive knowledge transfers to agentic tasks with zero fine-tuning
📑 Paper: arxiv.org/abs/2606.24597
📖 Blog: qwen.ai/blog?id=qwen-agentwo…
💻 GitHub: github.com/QwenLM/Qwen-Agent…
🤗 HuggingFace: huggingface.co/collections/Q…
🧩 ModelScope: modelscope.cn/collections/Qw…
Rex Ma retweeted
New Science Blog: Why has AI advanced faster in coding than in biology?
To agents, bio databases are like cities built before cars—maddening to drive in because they're designed for different traffic.
How do we build infrastructure agents can use?
anthropic.com/research/agent…
Rex Ma retweeted
We are releasing Carbon: a crazy fast DNA model
Carbon is 275x faster than the next best model. So fast you can process the whole human genome on a single GPU in <2 days.
Here are the tricks we used:
When modelling DNA sequences a lot of the performance comes down to tokenizing the sequences in a smart way. BPE tokenizer struggle because there are no whitespaces and character (called base in DNA) level tokenizers waste a lot of compute on too many tokens.
Carbon is built with a unique tokenizer: we split sequences in chunks of 6 bases, but during both training and inference we can work with single base resolution. That's similar to having word tokens but resolving them at the character level. All possible thanks to the DNA tokens unique structure.
The architecture combined with the tokenizer makes the model 275x faster than the previous SoTA (Evo2) at this size.
We built an interactive demo so you can explore how the model can generate DNA sequences, investigate the structure of genes, predict the effect of mutations, generate and fold proteins and even reconstruct parts of the tree of life.
huggingface.co/spaces/Huggin…
