Insulate against model availability across regions and cloud providers
Insulate against model availability in different regions or different cloud providers.
46,249 results · Most relevant
For understanding everyday Spanish through listening and conversation.
Insulate against model availability in different regions or different cloud providers.
When you want to improve a prompt, run reflection with an LLM or agent that reads the collected feedback and proposes a better prompt, obtaining the score along with textual feedback that can be very domain-specific, rather than just a reward score.
Lakshya A. Agrawal advises that if your domain produces expert feedback, compiler error messages, profiler messages, or tool call error messages, you should return that as side information to the optimizer / fitness function.
Roman encourages founders to focus on making something that is not possible now possible; users will come to use that thing, and by pulling them to the next impossible frontier a company gains distribution and data advantages.
Roman attributes Cursor's success to how quickly the team adjusts to the reality of the market, and recommends building the best-in-class experience for a thing that is going to happen, even if that thing exists elsewhere today.
Roman Ugarte recommends that when having a product debate, you ask 'How would a human do this? What would you want from your teammates in this situation?' He says the answer is often clarifying and pretty unanimous, and then you just need to build it. He adds that this doesn't require being a genius; you just need to ask the question and push AI to behave similarly.
Roman recommends maintaining the 'startup feeling' inside the company as it scales: everyone moves fast, deeply trusts each other to execute their part, and is fired up about a clear vision they know they need to execute on. He considers this critical for continued success.
hedged
Elie Bakouch recommends building a multi-agent discovery system with generators, a judge, and a speedrun reward loop instead of pure evaluation. He suggests including both closed and open source models in the generator pool because open source models are super effective for the cost, and giving the judge a sense of taste about whether a method is good, outside the loop. Humans are also super useful to judge agent ideas and steer them in the right direction.
Elie Bakouch recommends building open evaluation environments for AI research capabilities so the community can independently verify lab claims, and argues that part of recursive self-improvement research should happen in the open since many people working in AI are not at big labs. He also stresses making it easy for people outside big labs to understand how these models work and do research.
hedged
The creator advises that there's always room for face pulls regardless of whether you weight train two, three, or five days a week. On a two-day split, do one corrective and one hypertrophy session, with the corrective at the end of the week. On a three-day split, put the corrective at the end of the week and the hypertrophy session in the middle. On a five-day split: hypertrophy first, corrective after, hypertrophy after the day off, then one more corrective — keeping the balance in favor of corrective sessions. On an odd number of sessions, do more corrective sessions to get the neuromuscular re-education.
The creator recommends getting a long rope of around two feet as a good investment that allows doing the face pull properly; if unavailable, set the machine up with two traditional ropes and pull them all the way to their end. A long rope allows getting the elbows in the right position without forcing yourself into the wrong one, while a short rope restricts how wide the elbows and hands can get.
Jemima Abu advises making it mandatory for your agents to follow your accessibility guidelines: every single time your agent runs, tell it to make sure you are building things in a way that people can use them, not just the way you think they should be used. It's easy to forget about accessibility when you have three hours to the deadline, but if your agent has it in mind, it's more likely to actually get done.
Jemima Abu advises that when prompting your agent, you should tell it to default to the best way of writing things. She notes that for web products there is a popular way to do things, but it's not always the right way, and says it's the developer's job to tell AI: don't go for what's popular, go for what is best. Part of your skill should also look out for cases where AI defaults to the popular rather than the best implementation.
Recovery mechanisms should not be applied uniformly to all failures — not every failure can be recovered the same way.
As agents are given more autonomy, they should also be given more reliable ways to recover from their actions.
Different recovery strategies are needed depending on the failure phase.
Vinoo Ganesh says we are now out of the search phase of AI and need to be building infrastructure, scaffolding, harnesses, and substrates that enable us to build a database to make use of this technology.
Vinoo Ganesh recommends making sure you have a good mechanism of actually constraining the set of documents that come back with each query, which he frames as really important for protecting alpha.
The way to solve all these security problems is scoping searches with security built in.
Nouha Dziri recommends framing team design as a compute allocation problem: starting from a solo agent, ask where the next dollar of inference buys the most reward. Teams of agents cost more than a solo agent, so the question is where the next dollar of inference-time compute should go to maximize reward at lower cost.
Choose sources → to get a library of your own and a daily brief.