Key Takeaways
- Software agents connected directly to Coinbase accounts are driving automated crypto trading volume growth of roughly 20 percent week over week.
- Specialized open-weight models organized in swarms routinely outperform large generalist foundation models on narrow enterprise tasks.
- Every line of code written at Coinbase undergoes automated inspection by dedicated cyber expert models before pushing to production.
- Armstrong's biotech company, NewLimit, plans to begin its first human clinical trials for epigenetic cellular reprogramming next year.
- NewLimit moves candidate longevity therapies from computer simulations to living patients through Armstrong's In-Silico to Clinical Trials Pipeline.
The Armstrong's In-Silico to Clinical Trials Pipeline
In drug development and longevity science, software predictions mean very little until tested against physical cells. Armstrong explained NewLimit's operational funnel from computational models to patient treatments: “We start with AI at the top to test millions of hypotheses... We take the most promising ones and go through our wet lab in high throughput screens.” (play from 2:25:46) The sequence requires four distinct gates:
- Step 1: Computational Hypothesis Testing (In-Silico): Start with AI models at the computational layer to evaluate millions of hypothetical therapeutic targets and cellular reprogramming interventions.
- Step 2: High-Throughput Wet Lab Screens: Take the most promising computational hypotheses into a physical wet lab using automated high-throughput screening to test functional cellular rejuvenation and epigenetic reprogramming.
- Step 3: Functional Assays and Preclinical Animal Models: Filter the top hits from wet lab screens and validate them through rigorous functional assays and mouse disease models to establish safety and restorative efficacy.
- Step 4: Regulated Human Clinical Trials: Advance validated candidates into phased human clinical trials to generate human safety and efficacy data required for formal regulatory approval.
When This Works (and When It Doesn't)
This pipeline works when you operate automated screening hardware capable of testing thousands of biological samples every week. Computational models strip away millions of dead ends at near-zero marginal cost, allowing researchers to allocate physical wet lab capacity to targets with the highest probability of success. It creates a closed loop where physical assay data continuously refines the computational model.
It breaks down when teams assume software simulations can bypass biological reality. Code cannot fully predict off-target mutations, tissue toxicities, or systemic immune responses in living animals. If physical assay cycles take months instead of days, the computational models train on obsolete data and produce unviable molecules. NewLimit still validates its top hits in mouse models before moving to human testing because software alone cannot prove that a cell has regained youth.
What to Do With This
Treat software models as top-of-funnel filters, not final decision makers. If you build automated software or data-driven products this week, place hard operational guardrails between your models and your production environment.
Armstrong advocates building swarms of narrow, specialized models rather than relying on one general foundation model. At Coinbase, dedicated security models review all code before deployment: “every line of code that we write at Coinbase before it goes to production gets reviewed by one of these cyber expert models and so we can look for any kind of issue or vulnerability upfront.” (play from 2:24:08) Apply that setup to your engineering workflow by deploying a fine-tuned, open-weight model to audit pull requests for security flaws before human review. Restrict autonomous agents from holding production admin keys or unrestricted account access until they pass automated verification.