35,595 results · Most relevant
Thorsten endorses adopting WebAssembly on the server side as a solution to the frictions of region-bound architectures, describing it as a new way to build back-end applications that is being done successfully. He offers the analogy that WebAssembly relates to containers as containers relate to virtual machines: binaries that are super tiny, easy to distribute, super fast to run, and portable.
The creator argues that getting the workload off your machine and into CI is key for optimizing costs: without visibility into how agents work with the tool, you can't tell whether they're too verbose, achieving the task, when tool calls happen, or whether they succeed — and you have no opportunity to optimize. Once agents are in the pipeline, cost optimization, A-B testing of models, and performance analysis become possible.
Working locally on a single session or device means a device failure or network loss interrupts the workflow; the creator argues agents should run in the platform so they run autonomously and are not interrupted by local device or network issues.
Treating tools as the full context is a mistake. A context graph helps you reason across the whole company and must include data, synced data, company skills, and processes — not just tools. Relying on tools alone yields an incomplete graph that cannot reason across the company.
AlphaFold 2's design was not accidental: it baked in scientific intuition from biophysics and biochemistry — that amino acid residues influence each other rather than acting independently. Baking inductive bias from scientific knowledge into a model makes it much more data efficient because it doesn't have to replicate everything from data. Sal adds that with smaller amounts of data, having more inductive bias in the model helps — you actually need it to get results at smaller scales of data.
explicit
Evals are not built once — they're a flywheel that evolves as your customer evolves. As customers move through the journey, what they try to do with your product changes and gets more complex, so evals must stay up to date and reflect exactly what the customer is trying to do right now. Evals must reflect what the customer cares about, not what you think they care about, so get customer input and make sure evals reflect their problem.
The solution to the benchmark illusion is not better benchmarks but closing the customer loop: get feedback from your customer and feed it back in. Success in the eval flywheel should be defined as what the customer thinks is success, not what you think is success.
In a world where anything is possible, deciding what to build is more important than ever. Agents cannot replace the human work of deciding what matters, understanding what is being built, interpreting ambiguity, owning the system for the long haul, exercising judgment and restraint at the architecture level, and knowing what not to build.
Bad engineers who don't review details, don't refine the spec, and aren't detail-oriented end up producing more, faster — while detail-oriented engineers seem slower. This dynamic yields performance anxiety that pressures engineers to merge PRs faster than they can judge, leading to code being merged without full understanding, and short-term thinking — getting the PR in — wins over long-term code health.
The goal is not just code generation; the goal is human knowledge, human understanding, and human review. We need to optimize for that human layer. AI does not remove the cognitive overhead tax; unless tamed, it exacerbates it. An agent can generate five plausible patterns in five different styles, forcing the human reviewer to reverse engineer not just the code but the intent behind it.
The cost of creating code is falling, but the cost of understanding a software program is not. If AI produces more than humans can possibly evaluate, it has not accelerated the team — it has created a bottleneck.
Software is a record of technical, product, operational, and user-facing choices. If engineers do not see those choices, they cannot own them — and without ownership they are not really engineering, just accepting output.
AI output arrives with the shape of completion, but the real work has just moved to the reviewer. What AI is really doing is producing more, faster — more characters that eat more human time.
The real danger of AI slop is that engineers never notice there was a choice to make; slop is about the absence of judgment, not about AI as a tool. When work arrives already smoothed over, the unknown unknowns stay unknown.
An agent operates much more effectively inside a codebase that has structure, patterns, clear boundaries, and rationale. But if the system is already a big ball of mud, the agent does not rescue you from the mud — it just builds the ball bigger, faster.
The guests argue that no one will zero-shot a room-temperature superconductor: big physical discoveries require actually running experiments. They stress distinguishing results in math and theoretical physics — where zero-shotting works differently — from the physical world, which is far harder. Optimization against answers known in papers or textbooks is not enough when going beyond existing knowledge, and optimization against math has high precision, unlike physical lab work with its variants, uncertainties, and aberrant measurements.
Machine learning only works on what it has been trained on, but scientific discovery is by definition what you haven't been trained on. You cannot think your way to new knowledge: the universe is too complicated, and no amount of rereading textbooks lets you anticipate all experimental outcomes. Progress requires an iterative loop of conjecture and experimental testing, which is why the team is building physical labs. They add that intelligence is necessary but not sufficient for creating new knowledge — new knowledge is created when ideas are found to be consistent with reality.
Tara Sesha argues that in the current AI era, urgency in getting functionality out matters more than getting it perfect a priori; she shifted her own mindset from perfection to shipping and iterating as fast as possible.
endorsed
Nan Yu argues that if you take users on the journey with you, they are far more accepting of change, churn, and product evolution — especially when they have transparency into what's going on behind the scenes. As long as there is a coherent story users can follow, there is a lot you can do without burning bridges.
endorsed
Nan Yu observes that the most successful agent design efforts have tried to map onto human nature: we have millions of years of evolution teaching us how to hold things in our heads, and 40 agents is quite a lot — so group agents into bundles of activity rather than managing many individual agents.
endorsed
Choose sources → to get a library of your own and a daily brief.