TAI #219: AI, Cancer and the Future of Personalized Medicine

Also, Nvidia’s AVO Agent, GPT Sol 20% price cut, Deepseek’s multimodal V4-flash and more

What happened this week in AI by Louie

On August 15, Dario Amodei argued that AI companies will earn trust by curing diseases, rather than promising they might. Anthropic is increasing its work in biology and medicine, and Amodei believes most human disease could be cured within five to ten years. That is an ambitious forecast, but the past week offered a clear picture of why the idea deserves serious attention.

Anthropic showed AI agents coordinating the design of new proteins and having their work validated in real laboratories. Moderna and Merck reported that a personalized cancer therapy designed with machine learning succeeded in a large Phase 3 trial. Other researchers demonstrated faster automated protein testing, engineered cancer-targeting immune cells, and new systems for simulating how cells respond to treatment. These developments cover different parts of the same problem: specialized models identify targets and design molecules, LLM agents coordinate research, automated laboratories test the results, and clinical trials establish whether treatments help patients. What is changing is how quickly those parts are starting to connect.

The clearest patient result came from Moderna and Merck. Their INTerpath-001 trial enrolled 1,137 people whose high-risk melanoma had been surgically removed. Some received Keytruda, an established immunotherapy. Others received Keytruda plus a personalized messenger RNA treatment called intismeran autogene. The combination improved recurrence-free survival and reduced the risk of the cancer spreading to distant sites. This therapeutic cancer vaccine is designed to stop an existing cancer from returning, with each treatment built around the mutations in one patient’s tumor.

The process starts by sequencing the tumor and the patient’s healthy cells. Comparing the two reveals mutations unique to the cancer. Researchers then analyze tumor RNA to determine which mutated genes are active. Some produce altered protein fragments, called neoantigens, that can help the immune system recognize cancer cells. Choosing the right fragments requires accounting for which mutations the tumor expresses, which fragments the patient’s cells can display, and which are most likely to trigger an immune response. Moderna’s system scores those possibilities, selects up to 34 targets, and combines them into a single personalized mRNA sequence.

That sequence is packaged into a lipid nanoparticle and injected. The patient’s cells produce the selected fragments, training immune cells to recognize cancer cells carrying the same mutations. Keytruda helps sustain that immune response. Targeting dozens of mutations also reduces the chance that surviving cancer cells can escape by losing one target.

Moderna says the design process runs from raw sequencing data to a patient-specific treatment without manual intervention. Another software system coordinates manufacturing, quality control, shipping, and scheduling. The company began applying machine learning to mRNA design in 2014 and built its individualized-treatment algorithm in 2016, giving it a decade of purpose-built scientific AI. One clarification: the widely shared 49% reduction in recurrence or death came from an earlier 157-person trial. The new Phase 3 study met its endpoints, but its full results have not yet been published.

Anthropic’s research points to a different role for AI. Its models received an expert-written research protocol, access to scientific literature, and specialist protein-design tools. They then selected molecular targets, proposed protein structures, checked which designs should fold properly, and decided which candidates to send for laboratory testing. Of 1,320 designs that produced usable measurements, 354 bound to their intended targets, a 26.8% success rate. The agents produced at least one successful protein for 14 of 15 targets. Human scientists provided the protocol and laboratory validation, while the models coordinated a complicated scientific workflow with less hands-on research time.

A separate study showed how these designs can move toward treatments. Researchers used AI-designed proteins to engineer immune cells that recognize cancer-associated targets. Several early designs needed further adjustment before they worked properly inside living cells, but the revised cells went on to kill tumor cells selectively in laboratory experiments.

A blinded antibody benchmark also exposed a weakness in molecular design: AI systems often struggled to identify their strongest candidates. Their selections beat a comparison antibody only 9.8% to 13.8% of the time, while randomly chosen candidates did so 39% of the time. That makes automated experimentation more valuable. If a model can propose hundreds of plausible molecules but cannot reliably pick the best one, researchers can test more candidates, learn from the results, and improve the next round. David Baker’s group described a semi-automated system that can produce and characterize hundreds of proteins per day while cutting gene-synthesis costs roughly fivefold.

A strong example of where this is heading comes from OpenAI and Ginkgo Bioworks. They connected GPT-5 to Ginkgo’s robotic cloud laboratory, where software can order experiments, machines execute them, and results return to the model. GPT-5 designed experiments to improve cell-free protein production, analyzed the measurements, and proposed the next batch. Across six rounds, the system tested more than 36,000 experimental conditions on 580 automated plates and cut protein-production costs by 40%. Each round supplied the evidence GPT-5 needed to form better hypotheses and improve its next experiments.

This is just one part of a growing physical research infrastructure. Emerald Cloud Lab offers remotely controlled laboratory experiments. Recursion says its automated labs can process up to 2.2 million samples a week. Insitro combines more than 20 petabytes of automated cellular experiments with human genetic data. Vivodyne uses robotic systems to grow human tissues and test how they respond to drugs and genetic changes.

Virtual-cell models add another layer. GenBio’s AIDO Cell can simulate how cells respond to combinations of drugs or genetic interventions, while the Arc Institute’s new Virtual Cell Challenge asks models to predict how unfamiliar cells will respond when particular genes are suppressed. Better simulations can narrow the search; automated laboratories can test whether those predictions hold up.

---

---

Why should you care?

Dario is hoping AI cures are a key route to making AI more popular. I think AI medicine could also become one of the main ways AI gets smarter. Scientific research offers frontier AI labs unusually valuable training signals: outcomes verified by the physical world. A protein binds to its target or it does not. A treatment changes a cell’s behavior or it does not. Those measurable results can drive reinforcement learning. An AI agent can study scientific papers, propose experiments, assess the results, revise its hypotheses, and try again. Each experiment creates fresh evidence about cause and effect, improving the model’s reasoning while producing useful medical discoveries.

I expect leading AI companies to spend hundreds of millions of dollars training agents this way - designing drugs and treatments becomes part of model training costs. The commercial incentive is strong: better scientific agents improve general reasoning while generating proprietary biological data, valuable intellectual property, and potential treatments.

But there is a big gap between the feedback an AI agent can obtain quickly and the evidence needed to prove a treatment works in patients. Automated laboratories can measure protein binding, immune responses, or changes in cell behavior within hours. Simulations can run even faster. Neither can tell us whether a patient’s cancer will return four years after treatment or whether serious side effects will emerge over time. Answering those questions requires clinical trials, long-term monitoring, and regulatory review. Until a treatment is approved and reaches more patients, the clinical data needed to improve future versions also accumulates slowly.

That delay will remain a real constraint, but better experiments and predictive models can still make clinical development faster and more productive. Stronger laboratory evidence, clearer biological mechanisms, and better patient selection should help more companies move promising treatments straight into combined Phase 2/3 trial. Regulation will need to keep pace with faster discovery and more personalized therapies while maintaining high standards for long-term safety and effectiveness.

Biology is also generating far more information than human researchers can interpret alone. Whole-genome sequencing, tumor DNA in blood, protein measurements, medical imaging, and patient histories each capture part of the same disease. AI can combine those signals, identify patterns across large populations, and apply them to an individual patient. Earlier cancer detection is one of the biggest opportunities. In GRAIL’s 35,878-person PATHFINDER 2 study, adding a blood test to standard screening increased cancer detection 6.5-fold, and 71% of newly detected cancers were at stages I to III. Better blood tests, sequencing, and protein analysis will make more cancers visible before they spread to distant organs.

That shift changes the incentives for treatment. A cancer found early has fewer malignant cells and fewer opportunities to develop resistance. Surgery, personalized vaccines, or targeted immune treatments have a better chance of eliminating it. Earlier detection creates a larger market for early intervention. Better early treatments make screening more valuable. Together, they mean more cures and fewer years managing advanced disease.

The treatment platforms are also becoming more adaptable. Moderna’s mRNA approach can use the same delivery system while changing the genetic instructions for each patient. CRISPR offers a related form of flexibility: the editing machinery can be directed toward a particular genetic mutation. Last year, researchers at Children’s Hospital of Philadelphia treated an infant with a customized CRISPR therapy designed for his rare genetic disease. The obstacle has been the scientific labor needed to customize these treatments safely and cost efficiently, since each patient may require different molecular targets, treatment instructions, and tests. AI agents and automated laboratories can supply the research capacity, while better software and manufacturing systems improve quality control and patient safety.

Amodei’s timeline is bold. His broader direction looks right to me. AI can help us understand disease earlier, design treatments around individual patients, test those treatments more quickly, and learn from every result. When the best AI companies have a direct incentive to train increasingly capable agents on real scientific discovery, progress in medicine could arrive much faster than most people expect.

— Louie Peters — Towards AI Co-founder and CEO

---

Hottest News

1. NVIDIA’s AVO Agent Hits a Perfect Score on ARC-AGI-3 Without a New Model

NVIDIA’s AVO agent completed all 183 levels across ARC-AGI-3’s 25 public environments using Claude Opus 5, earning a 100.00 Relative Human Action Efficiency score. AVO adds persistent memory, a supervisor that redirects stalled trajectories, and its own execution loop around the model. It completed the set in 6,624 environment actions, about 12% fewer than VISTA’s reported Opus 5 run, although NVIDIA stresses that the systems differ enough that this is not a controlled comparison. The same AVO architecture previously ran autonomously for seven days optimizing GPU kernels, exploring more than 500 directions and producing multihead-attention kernels up to 10.5% faster than FlashAttention-4 on NVIDIA’s tested DGX B200 configurations. The 100.00 result covers the public set only, not ARC-AGI-3’s semi-private or private competition sets.

2. OpenAI Cuts Frontier API Pricing by More Than 20% for Three Months

OpenAI temporarily cut GPT-5.6 Sol’s standard short-context API price from $5 to $4 per million input tokens and from $30 to $20 per million output tokens, reductions of 20% and 33% respectively. Cached input now costs $0.40 per million tokens, while prompts over 272K tokens are priced at $8 input and $30 output. The promotional rates are guaranteed at least through November 21, 2026, with no new model ID required. The reduction is also rolling into eligible ChatGPT Work and Codex credit pricing, while Plus, Pro, and Business subscription usage remains unchanged. Batch and Flex processing are cheaper again at $2/$10, while Fast mode remains the premium option at $8/$40. This follows July’s much larger price cut for Luna and a 20% reduction for Terra, leaving OpenAI’s full GPT-5.6 family substantially cheaper than at launch.

3. Claude Designed Protein Binders That Worked in the Lab for 14 of 15 Targets

Anthropic tested whether Claude could autonomously run the computational side of a protein-design campaign, and external labs found working binders across 14 of the 15 targets tested. Using Mythos Preview and Opus 4.8, Claude chose binding sites, orchestrated specialist protein-design and folding models, optimized candidates, and screened them before Adaptyv Bio and Twist Bioscience produced and tested the designs. Across all experimental arms, 354 of 1,320 designs bound successfully. In the 48-hour multi-target setup, Opus 4.8 achieved a 22.6% hit rate and Mythos 26.7%; giving Mythos a separate 24-hour run for each target raised its hit rate to 35.1%, versus the 10–15% Anthropic says is typical today. At least six targets produced high-affinity binders, and four matched or exceeded the best previously reported affinity. The campaign still relied on specialist protein models, substantial GPU compute, and weeks of wet-lab validation, so Claude’s role was to autonomously orchestrate the computational design process rather than replace the experimental work.

4. DeepSeek Unveiled an Experimental Multimodal Model

DeepSeek released DeepSeek-V4-Flash-Vision-Exp, an experimental multimodal version of V4-Flash that adds image understanding while retaining comparable text, reasoning, and agent capabilities. On DeepSeek’s own evaluations, it scores 36.5 on multimodal ApexBench versus 26.2 for text-only V4-Flash, 27.3 on Agents’ Last Exam, 64.3 on Chartography, and 35.0 on ZeroBench, putting its reported multimodal agent performance close to Claude Opus 4.8. The model also improves several text-agent scores, including DeepSWE from 54.4 to 59.3, although these results were run through DeepSeek’s own Harness configuration and have not yet been independently reproduced. Images cost no pricing premium: each is billed as at most 384 input tokens at regular V4-Flash rates. The API accepts mixed text and images through Chat Completions, Anthropic-compatible Messages, and Responses, while a new free Files API lets developers upload an image once and reuse it across requests. DeepSeek Harness 0.1.1 shipped alongside it with native support.

5. Cursor Launches Origin

Cursor launched Origin, moving beyond the editor to host repositories and pull requests directly inside Cursor. Developers can create Cursor-hosted repos or sync existing GitHub repositories in real time; GitHub remains the source of truth for synced projects, with pull-request comments, reviews, and merges reflected in both systems. The larger shift is bringing agents into the same surface as the repository: while browsing code or a PR, users can ask Cursor questions, have it make changes, update the PR, or push a branch. Origin also launches with integrations for Vercel, Depot, and Buildkite for preview deployments and CI, including support for existing GitHub Actions workflows. It is rolling out in early beta across paid plans, except enterprise organizations whose admins opt out, with Cursor saying more agent-native hosting features are still to come.

6. Anthropic Brings Claude Mythos 5 to Claude Security

Anthropic is widening access to Mythos 5’s cybersecurity capabilities without giving users direct access to the model. Claude Security, currently in public beta for Enterprise customers, can now use Mythos 5 to scan repositories for vulnerabilities and return each finding with a CWE category, confidence, and severity ratings, and a suggested fix; users can then open Claude Code to implement the patch, with human approval still required. Mythos itself remains isolated within the scan, so access through Claude Security does not allow users to prompt it for offensive tasks elsewhere. Anthropic is taking the same approach with cybersecurity partners, embedding Mythos behind purpose-built tools that expose specific outputs rather than the raw model. It also announced $35 million in Claude credits for open-source security through a new Defender Advantage Fund and plans to expand its vetted Cyber Verification Program, with broader Opus and Sonnet capabilities first and Mythos-level access to follow.

7. Ornith-1.5 Ships a Self-Improving Open Model Family

Ornith released three MIT-licensed Ornith-1.5 models at 397B MoE, 35B MoE with 3B active parameters, and 9B dense scales, extending its earlier “self-scaffolding” approach into a broader self-improvement training loop. Instead of training only on fixed human-created tasks and harnesses, the system proposes progressively harder tasks based on what the model has already solved, generates the tools and decomposition strategy needed to tackle them, produces solution rollouts, and uses reinforcement learning to improve all three stages together. Ornith reports its 397B model reaching 86.1 on Terminal-Bench 2.1 and 56.0 on DeepSWE, around Claude Opus 4.8’s 85.0 and 59.0 in its evaluation setup. The 35B model achieves 79.0 on SWE-bench Verified despite activating only 3B parameters per token, while the 9B model achieves 70.6 and is also available in quantized mobile formats. The “self-improving” label refers to this training process, not a deployed model continuously changing its own weights. Weights, including GGUF, FP8, NVFP4, and MLX variants, are available on Hugging Face.

---

AI Tip of the Day

A student in our 10-Hour LLM Fundamentals Video Course asked us a useful question: How many times should you run an important test before trusting the result?

For stochastic LLM tests, our practical baseline is five runs.

Suppose a difficult test passes 90% of the time. If you run it once and it passes, you might conclude that everything is working. Run the same test five times, though, and there is about a 41% chance you will see at least one failure.

For important tests, run the same case five times and record two things:

• The pass rate across all runs

• Whether it passed all five times

Keep the prompt, model, temperature, and other randomness settings, as well as the source context, fixed. Otherwise, you are changing the test while trying to measure its consistency.

Five runs is not a statistical guarantee. It is a simple way to catch intermittent failures that a single successful run can easily hide.

---

Five 5-minute reads/videos to keep you learning

1. Is More Context Making Your AI Agent Worse? Here’s the Fix Nobody Talks About

The article explains why larger context windows do not guarantee stronger agent performance by breaking the LLM call into six parts: user messages, system prompts, tool schemas, retrieved resources, assistant history, and tool call records. It notes that maxing out even a million-token window rarely helps, since results often peak near 65 percent utilization and shares four fixes: sharpening system prompts, writing precise tool schemas, retrieving resources selectively, and steering long-running, iterative agents through compaction, memory stores, and specialized, targeted sub-agents.

2. 3 Core Ideas That Make Attention In Transformers Easy to Understand

This piece breaks scaled dot-product attention into three core ideas: embeddings as vectors transformed through matrix multiplication, weighted averaging on the query-key-value database analogy, and learned projection matrices that discover contextual embeddings during training. It walks through computing attention weights and outputs step by step, closing with a clear explanation of why query vectors function as questions one token asks about its neighbors.

3. Context Engineering: The Discipline That Quietly Replaced Prompt Engineering

The piece explains the four failure modes of context engineering: poisoning, distraction, confusion, and clash, and explains the KV-cache economics behind stable prefixes and append-only history. It also maps six production techniques (curation, compaction, external memory through files, recitation, just-in-time retrieval, and sub-agent isolation) using Claude Code and Manus as practical examples.

4. Stop Using Long-Context Windows for AI Agents (Build This Instead)

Long-context windows cost 15 times more than persistent memory retrieval, and this article argues for replacing brute-force context with a four-tier memory architecture: working, episodic, semantic, and procedural. It compares Mem0, Letta, Zep/Graphiti, and AWS AgentCore across storage, retrieval, and injection pipelines, then tackles decay policies using Ebbinghaus-style forgetting curves. It also flags production risks often skipped elsewhere: memory poisoning, drift, and hallucinated facts.

5. Claude Cowork Is for Engineers Too — Here’s How to Wire It Safely

This article draws a sharp line between Claude Cowork and Claude Code and argues that engineers dismissed Cowork by mistaking office-worker marketing for a limitation of capability. Cowork integrates with Slack, Jira, Drive, and Confluence via skills and connectors, handling spec drafting and incident triage that a terminal agent cannot touch. It also details failures, including permanent file deletions and a VM sandbox escape, which prompted Anthropic to default to cloud sessions.

Repositories & Tools

1. ECC adds a structured engineering workflow to coding agents, with planning, testing, review, persistent memory, reusable skills, and security scanning for Claude Code, Codex, and other agent harnesses.

2. OpenHuman is a local-first personal AI system that builds persistent memory from your data, orchestrates durable multi-agent workflows, and researches across your connected sources and the web.

3. FreeToken is an edge-native MoE serving engine that combines GPU, CPU, and system memory to run frontier-scale open-weight models, including 290B+ models, on consumer hardware.

4. Is Agentic is a website agent-readiness scanner that scores how easily AI agents can discover, retrieve, understand, and interact with a site’s public content and interfaces.

5. TensorRT-Model-Connect builds TensorRT engines directly from supported Hugging Face or local checkpoints, packages them into portable bundles, and exposes them through native C++ APIs.

Top Papers of The Week

1. FreeToken: Efficient Edge-Native MoE Serving with Bandwidth-Adaptive Execution

Frontier open-weight MoE models assume datacenter infrastructure, but FreeToken treats a personal machine as a unified, elastic inference platform. Rather than committing to a fixed offloading strategy, it continuously maps computation and model state onto whatever resources are actually available, co-designing model layout, expert residency, CPU-GPU execution, and memory management around each machine’s specific hardware balance. It scales from a 35B model on a laptop with an 8GB GPU to the 753B GLM-5.2 on a single workstation GPU, delivering 1.5–2.3x higher decode throughput than existing edge serving systems.

2. StateM: Reaching 95.3% Raw Accuracy, or a $15 Frontier Run, on Terminal-Bench 2.1 via Harness Scaling

Long-horizon agents fail even when their underlying models can solve the constituent steps: they lose track of mutable state, skip known procedures, or stop prematurely. StateM is an agent-native runtime that organizes execution around durable states, phase-local context, checked transitions, and recoverable runbooks. Without changing model weights, it raises GPT-5.5 xhigh from 83.1% to 92.1% on Terminal-Bench 2.1, surpassing GPT-5.6 Sol Ultra (91.9%). The same runbook transfers unchanged to GPT-5.6, reaching 95.3% raw accuracy across 445 trials at approximately $15 per full benchmark run.

3. Agent Lightning v1.0: Towards Harnessed Agentic RL

In harnessed agentic RL, the deploy-time harness owns the environment interaction loop while the trainer observes only LLM request-response pairs. This differs fundamentally from traditional agentic RL and introduces challenges in retokenization, sample merging, advantage calculation, and loss normalization that existing frameworks handle inconsistently. Microsoft’s Agent Lightning v1.0 is a lightweight framework (approximately 3,500 lines) that addresses these challenges and supports arbitrary agent harnesses as a practical testbed for studying harnessed RL. Its disaggregated architecture, connecting agents to training through an LLM endpoint proxy, has since been adopted by verl Uni-Agent, AReaL 2.0, slime, and Polar.

4. Agentic ESOpt: Fine-Tuning Long-Horizon LLM Agents with Minimal GPU Requirements

RL struggles with long-horizon agent fine-tuning: backpropagation demands heavy GPU memory, and longer trajectories make credit assignment harder. This paper argues evolution strategies (ES) are a better fit. ES enables full-parameter optimization using only inference-level GPU memory, performs trajectory-level parameter attribution without decomposing credit across individual steps, and composes naturally with prompt-space optimization like skill evolution and test-time search. Agentic ESOpt samples full-parameter perturbations, evaluates the resulting agents with environment rewards, and applies online reward-weighted updates, enabling on-the-fly parameter adaptation within prompt-space optimization loops.

Quick Links

1. Harvey introduces Harvey Tenet, its first post-trained open-weight model, built with Fireworks Research on a Kimi K3 base for long-horizon legal work. Tenet completes almost twice as many held-out LAB tasks and 20% more LAB Contracts tasks than base Kimi K3, while reward shaping also reduces unnecessary tool use and tokens. It reaches state-of-the-art performance on LAB Contracts, ranks second overall on LAB, and transfers its gains to unseen benchmarks, including APEX Agents and Redline Bench.

2. Liquid AI releases LFM2.5-DSpark Draft Models, ~300M-parameter speculative decoders for its 1.2B, 2.6B, and 8B-A1B models. They let the smaller draft model propose tokens for the full model to verify in parallel, delivering up to 3.18x higher GPU throughput and 2.87x on-device speedups without changing greedy-decoding output quality. For LFM2.5–2.6B, Liquid also reports a 57% average reduction in function-calling latency, with support for llama.cpp and SGLang available at launch.

3. Cartesia released Sonic 3.6, its latest real-time text-to-speech model, with improved naturalness and expressiveness across 44 languages. Now in beta, it ranks #1 on both Artificial Analysis streaming speech leaderboards, including the controlled-voice test that evaluates models using the same reference voices rather than each provider’s best voice catalog. Cartesia says the update comes from fundamental model improvements informed by feedback from teams using Sonic 3.5.

Who’s Hiring in AI

Senior AI Engineer/Forward Deployed Engineer @Towards AI (London/Hybrid)

Forward-Deployed Engineer @Deepgram (Remote/USA)

Agentic AI Solutions Engineer — Banking @VAM Systems (Dubai, UAE)

Senior Forward Deployed AI Architect (GenAI, AWS) @Provectus (Remote)

Forward Deployed Engineer @Loadsmart (Remote/USA)

AI Engineering Lead @Blend360 (Hyderabad, India)

Staff Embedded Software Engineer @Northrop Grumman (Annapolis, MD, USA)

Interested in sharing a job opportunity here? Contact sponsors@towardsai.net.

Think a friend would enjoy this too? Share the newsletter and let them join the conversation.