High Agency and Autonomy Over Policing
Grant employees high agency and autonomy to execute without excessive policing or fear-based friction. This trust enables speed and allows employees to move quickly, make mistakes, and learn.
high
35,595 results · Most relevant
Grant employees high agency and autonomy to execute without excessive policing or fear-based friction. This trust enables speed and allows employees to move quickly, make mistakes, and learn.
high
Trust is the foundational value that enables high agency, rapid experimentation, and hiring quality. Leaders prioritize holding onto trust as the company grows to maintain speed and avoid bureaucratic bottlenecks.
high
Treat AI agents as untrusted entities with oracle-like capabilities when designing threat models. The untrusted side in a threat model is assumed to be an oracle. The question is whether information can be extracted from the trusted side.
The early internet was plagued by viruses and security failures, but society continued building it and learned to make it safer. This historical precedent suggests AI should follow a similar path of iterative security improvement rather than halting development.
Pacing AI development is logically impossible without a published schedule or known baseline rate of development. Claiming to be 'pacing' without a baseline is disingenuous. If researchers genuinely believe in existential risk, they should nationalize the technology; if not, it is an HR/retention issue.
Current access and security models are not granular or performant enough to handle AI agent swarms. A different access and security model is needed going forward, leveraging AI to handle the usability complexity of multi-layer security.
Zach Lieberman states that he has a group of graduate students thinking about future tools for artists and designers. He teaches them about art and wants them to engage with history. He advises students to be like a sailor, using the wind of hype to go in their own direction, and to engage with the tools and voice of today.
Ofir Ehrlich and Gonen Stein argue that a company's proprietary data is its primary competitive advantage, more valuable than models or compute, which have near-zero switching costs. This data, often locked away, has become a strategic moat.
Gonen Stein and Ofir Ehrlich state that high-quality, non-synthetic data reflecting real-world complexity is scarce. To build significant AI applications, one must interact with real-world companies and their data, as synthetic data often fails to represent real-world complexity.
Max Hodak endorses the view that the brain is literally a computer that solves computational problems by arranging matter, and that treating it as such enables large effect sizes in medicine that drugs cannot match.
Hodak reports that Science Corporation received marketing approval in Europe for Prima in July, making it commercially available, with first sales expected soon.
Hodak states that the central objective of BCI is to restore visual, auditory, balance, and motor signals in and out of the brain, which is an end in itself and the core of neuroscience.
Alex Hormozi states that the three non-technical reasons entrepreneurs fail are lack of focus, lack of patience, and low frustration tolerance. He argues that most entrepreneurs fail because they lack character, not tactics.
Alex Hormozi identifies three required character traits for business success: focus (doing one thing), patience (doing it for a long time), and frustration tolerance (enduring setbacks). He notes these traits are 'anti-human' because humans want novelty, immediate gratification, and ease.
The speaker advises that in the mid-term, production cannot wait — without the right amount of eval structure, organizations fail to deliver the promise in production that they make with the LLM. He cites the MIT Gen AI Divide report, which found 95% of organizations are not able to get any returns because they are not delivering the promise in production that they tried to make with the LLM.
Bakken's standing advice for people getting into running or starting to have trouble is to try modern running shoes with a carbon plate or modern foam. He also recommends switching shoe brands to see if a different brand benefits you more. He observes that since roughly 2018-19, carbon-plate and improved-foam shoes reduced running injuries and let more patients keep running in his clinical practice of about 1,200 patients.
Consequential creator advice requires independent verification before Today can promote it.
Bakken recommends doing more marathon-specific work with longer, more durable sessions from about 6 to 12 weeks before the marathon, limited to those marathon-specific weeks rather than year-round. Before that block, he says you can use the math from his book to build a very good base.
When adding intelligence to gateway connectivity, think about end-to-end user identity connected to the data; for orchestration, think about context assembly, memory, and context construction; at the memory-core/database layer, extend into ontology, semantic layers, and memory.
When information comes in, there are different types of memory, and you want to route it effectively and cheaply, which can be done with decision models; normally memory routing would be done using some code, but with decision models it becomes very easy and very cheap.
Zach Lloyd recommends building a private benchmark suite by taking any past run from your system and adding it to a benchmark set — starting by hand with a set of small, medium, and hard representative tasks. He also notes it is possible to have the factory itself look at prior sessions and identify interesting ones for the benchmark set, since everything is stored in code. He frames the guiding question as 'What's the best model for me to use for front end tasks?' and the curation framing as 'this is a representative thing that I did in my factory that could be useful to measure'.
Choose sources → to get a library of your own and a daily brief.