35,848 results · Most relevant
Test explosive power with a counter movement jump on force plates, include rotational power (e.g., via Proteus) for position players' swings, and use force-velocity profiling (BBT) to assess lower body strength for pitchers and position players.
Keep in-season lifts to 15-20 minutes so they fit as part of pre-game preparation before positional work and batting practice, using a dose of 3-5 exercises (or 4-6 depending on time of year) per session.
Consequential creator advice requires independent verification before Today can promote it.
Assess speed with a sprint assessment for position players, focusing on base running and stealing bags, and use tools like Vew Motion or the 1080 for sprint force-velocity profiling and kinematics outputs.
For securing AI itself, we won't be able to reason about these massive large language models, so the best option is something more like securing the boundaries; reasoning that a 500 billion parameter model isn't going to do something crazy is not really tractable.
If you are a bank, hospital, financial service, or government, you should be investing heavily, heavily, heavily into cyber defense.
Consequential creator advice requires independent verification before Today can promote it.
The speaker recommends that banks, hospitals, financial services, and governments invest heavily — heavily, heavily — into cyber defense, saying there are many things organizations can do to tighten up their ship. He notes retail systems are very antiquated and haven't been touched in a long time, and that simple methods such as trying admin-admin credentials actually get you into the system. Almost all social security numbers were breached because organizations didn't change the admin-admin password, so organizations should change default admin passwords to prevent breaches.
Consequential creator advice requires independent verification before Today can promote it.
When you want to tweak agent behavior based on the type of work, use per-session effort knobs and field overrides instead of looking for a dedicated knob; if you want the agent to take different types of action based on rough sizes of work, system-prompt that behavior rather than looking for a dedicated knob.
Drive Cloud managed agents and generally use the Cloud API from inside Claude Code: direct Claude Code to spin up resources and it uses the ant CLI to interact with the API; saying 'spin up a managed agent that does X' is a recommended way to jump in without the console. When Claude Code kicks off it makes API calls to create the agent, session, and resources.
Run dynamic workflows on schedules or listening to webhooks — a GitHub Bug Hunter workflow is a classic example; schedule a watchlist scanner five minutes before market open so the report arrives before open; trigger flows from Slack or other interfaces; and hook dynamic workflows up to CI.
Enable multi-agent sessions through the API by passing the multi-agent argument.
Use the Terraform provider to create agents programmatically for sharing across an organization; use export-style functionality to export agents and drop them into other places; when you want to share templated agents, upload the agent body or API resource to another place rather than expecting a one-click share button — there's not a first-class sharing mechanism within the API right now.
Use the hill climb skill (slash Claude API hill climb) in Claude Code: take an off-the-shelf prompt like the watchlist scan prompt, tell it what to iteratively improve, and it grinds until it gets a nice prompt. Claude may or may not give you a human interface to iterate on depending on the problem; you can request one.
partial
Pull the full transcripts of flow runs, give them to Claude, and ask it to identify where the flow is doing too much or not enough; you can also ask Claude to optimize flows based on criteria you care about like tokens or capability.
CJ Avilla recommends using Haiku for the scan phase — a fast, efficient model, maybe two Haiku agents per ticker — and investing more compute in review with a Fable agent, two or three per ticker, when details are being missed.
In hindsight, the speaker says waiting to go full-time was a mistake and recommends going full-time on the business sooner rather than later; when co-founders felt it was the right time to transition, the team supported them and kept going.
hindsight judgment with moderate confidence
Quigley recommends structuring a screening campaign so results become a database lookup, enabling a huge number of agent rollouts. He also advises including choices a human would not make — counterfactuals, counterintuitive decisions, mistakes — in RL training data so the model doesn't just learn chemist sociology fingerprints. Building a large-scale reinforcement rollout dataset from activity cliffs and protein variations works because each small molecule change has a known causal, reproducible effect, so agent trajectories double as reasoning traces at scale.
attributed to the guest as recommendations
Quigley advocates representing small molecules and proteins in readable, interpretable English instead of embeddings — natural language acts as a regularizer against the memorization problem. Leash taught a language model medicinal chemistry by having it describe structures in natural English and produce diffs between two molecules in natural English, before asking it to reason about binding. He argues natural-language reasoning agents are more resistant to memorization than embedding-based models like Hermes, because as parameter counts grow, even mutant-trained models will eventually memorize the mutants.
endorsed as Leash's approach and belief
Quigley advocates pursuing large amounts of diversity — more proteins and more diverse chemical libraries — as the way to scale chemistry models, including intentionally chosen activity cliffs. Leash engineers mutations in proteins to create fine-grained activity cliffs, teaching models to detect very small changes that drive outsized behavioral differences. He notes a tension: experiments must balance coarse diversity (finding binders across many protein targets) against fine-grained subtle changes (teaching models activity cliffs), and both matter when designing chemical material.
endorsed as Leash's strategy; the tension is described as a design consideration
Harish Gaggar recommends that teams in air-gapped environments build their own MCP servers rather than using outside-available ones, leveraging the FastMCP server package or wrapper to create different sets of tools inside the wrapper.
Harish Gaggar recommends using the LangGraph framework from the LangChain community to build task-specific nodes based on your questions, and plugging in marketplace plugins, skills, and prompts so agents respond efficiently.
Choose sources → to get a library of your own and a daily brief.