35,800 results · Most relevant
Jay Sahnan argues that any agent needing to act on the web should have the browser as a tool, since many sites lack APIs or MCPs, and that browser agents can access information behind login portals and antibot-protected websites.
Voss argues the knowledge of what a business actually needs isn't in the training data or the weights — it's in the mind of the person who runs that business, and the only way to get it out is by talking to them. That work doesn't transfer between domains: a calendar app and a scheduling app solve almost the same problem but are completely different software with totally different requirements and definitions of good.
Voss argues the profession needs to fix the junior engineer problem — the junior developer career ladder has disappeared before the industry has figured out how to train new juniors. Universities teach data structures and boot camps teach React, but nobody teaches 'go sit with a baker for a week and come back with a spec' — which is the person we need. There is a whole cohort of unemployed juniors, and the industry needs to figure out how to create product engineers on purpose instead of accidentally, and quickly, or everyone will be screwed.
Voss argues that writing down what good means is where the value is now, since anything you can verify cheaply you can train against until you beat it if you are a model maker.
Several model-level techniques reduce inference flops: Mixture of Experts with K experts reduces test-time flops over all experts by roughly one over K; SSMs avoid full attention on every layer (maybe only every fifth or tenth layer); multi-headed latent attention compresses in the sequence domain; quantization down to even 2-byte models reduces cost; and you can distill a big model into a smaller one when the smaller one is all you need.
Not every request should go to the biggest model; route requests to cheaper/smaller models where possible to save flops. Use a waterfall or tool-call approach that escalates to bigger models only when needed, which saves a lot of flops.
When mapping PyTorch ops to hardware ops, do not go all-to-all; insert a bottleneck layer in between to map down through. The RISC-style primitive op set serves as that bottleneck.
If a model's useful life is around two years, inference compute is roughly 12x more important than training compute. Francois recommends weighting inference flops about 12x more than training flops when selecting models for a paper, rather than only comparing ISO train flops. He notes this 12x figure depends on the assumed model useful life.
The 12x ratio depends on an assumed two-year model useful life, which Francois himself flags as uncertain.
Pre-fill and decode are very different workloads that should be served in different ways, potentially disaggregated onto different GPUs. Serving them separately raises questions about how to disaggregate them and whether different parallelism or kernels can make each more efficient. Disaggregation choices affect user experience and total GPU throughput.
The biggest challenge with routing across providers is that not everyone serves models the same way — which is a good thing because different teams have excelled at different performance optimizations. But one bad apple can ruin the bunch: if you have a single provider that's underperforming on one of the four criteria, you might have a customer experience that's the least common denominator. A provider may have over-quantized a model, be running it on faulty hardware, or be using a different version of the weights.
The first step when an inference request arrives is request-level routing: deciding which GPU replica should handle the request. Routing must balance load across GPUs AND route to wherever the KV cache is already resident — an art that must weigh both. In the old world of databases and web services, you could simply round-robin requests away from overloaded computers; in the LLM era, KV cache residency adds a constraint that breaks naive round-robin.
None stated; described as a core serving principle.
When routing across multiple inference providers, a single underperforming provider on cost, intelligence, speed, or region/privacy degrades the whole customer experience to the least common denominator. Evaluate inference providers across four dimensions: cost, intelligence, speed, and region/privacy. The definition of an inference provider is loose — the multi-provider quality problem also applies within a single provider with different deployments on different hardware, regions, teams, and inference engines.
None stated; presented as an endorsed heuristic from Jeff (Ollama CEO).
Dr. Scott Forbes states that believing creatine causes dehydration and increases hamstring pulls is a myth — evidence shows the complete opposite. Creatine actually reduces the incidence of muscle strains, muscle cramps, and muscle injuries over a football season in American football players (Rick Kreider study).
Dr. Scott Forbes states that based on current evidence, there is no evidence that creatine causes hair loss. A 12-week randomized controlled trial in Iranian participants found no effect of creatine on DHT levels, and a separate 12-week randomized controlled trial using direct measures from a hair specialist found no detrimental effect of creatine on hair health or hair quality. DHT (dihydrotestosterone) has been linked to male pattern baldness (alopecia), which is why the DHT question was examined.
Dr. Scott Forbes states that creatine is probably the most effective dietary supplement, particularly when combined with exercise and strength training, for making muscles and lean mass healthy and improving them.
hedged
Dr. Forbes explains the main theory of how creatine works: a larger creatine energy system lets you do one or two extra reps per set, and training harder over time makes you bigger and stronger.
presented as the main theory of how creatine works
Dr. Scott Forbes says you can take creatine daily for long periods of time, citing data showing it is safe for very long periods, with the longest study in the literature lasting just over 14 years.
Presented as what the literature shows
Dr. Forbes says creatine is probably the most effective dietary supplement, particularly when combined with exercise and strength training, for making muscles and lean mass healthy.
speaker hedges with 'probably'
Dr. Forbes and colleagues conducted a review quantifying creatine's effect on performance: the creatine group was 1.43 kg stronger on bench press and just over 5 kg stronger on squat versus placebo, and jumped about 1.48 cm higher. The average study duration was about 10 weeks. Both groups trained and improved; creatine adds a further benefit.
Some figures flagged as possible numeric errors in the source transcript
Dr. Scott Forbes notes creatine supplementation is safe based on the research. The longest reported creatine study lasted just over 14 years, showing long-term safety of daily creatine use.
Choose sources → to get a library of your own and a daily brief.