- Execution beats strategy every time. Good strategy is important, but execution is what matters. Focusing on execution allows us to iterate toward the strategy that wins for customers today, even as the world changes.
- AI will accelerate the race between technology and education. To thrive, focus on generating real, hard-to-replicate value, stay curious and keep learning. Choose the right race to run. (Career advice for young high-performers)
- People should avoid making pretentious expenditures (like buying a Rolex); they don’t bring much happiness.
- Being known via content/curation, etc., will not give you the greatest ROI. Just spend that time learning and doing and getting recognised for your work.
- Want to get into deep learning? Spend time understanding the fundaments and find interesting problems to solve (Spoiler: You don’t need a lot of GPUs)
- A few truths about LLMs a. LLMs can predict reasonable responses to any prompt b. You cannot know whether a given response is accurate c. You can estimate accuracy for a model and a given set of prompts using evals d. You can generally increase accuracy by using a larger model, but it’ll cost more and have higher latency e. Models generally get more accurate as the corpus it’s built from grows in size
- GraphQL makes many REST problems disappear. Clients breaking after API changes, waterfall requests, over-fetching data, endpoint discovery, security and performance.
- GetLago uses OLAP (online analytical processing) tools like Click-house to scale its events engine. They find OLAP tools are generally efficient at answering complex read queries, and good at batch processing many write queries.
- More major LLM releases happened this week (all under 24 hours!); Google released Gemini Pro 1.5, Open API released GPT-4 Turbo and Mixtral released Mixtral 8x22B
- Others
- Code search is hard
- Meta used monolithic architecture to ship the Threads app in five months
- The way to get luckier is to find what’s repeatable (or simply do what works)
- Intercom re-branded the company to be AI-first
Substack edition
Young people should avoid having pretentious expenditures (like buying a Rolex), there’s not much happiness in it
LLMs can predict reasonable responses to any prompt
You cannot know whether a given response is accurate
You can estimate accuracy for a model and a given set of prompts using evals
You can generally increase accuracy by using a larger model, but it’ll cost more and have higher latency
Models generally get more accurate as the corpus it’s built from grows in size
More major LLM releases happened this week (all under 24 hours!); Google released Gemini Pro 1.5, Open API released GPT-4 Turbo and Mixtral released Mixtral 8x22B
What to get into deep learning? Spend time understanding the fundaments and find interesting problems to solve (Spoiler: You don’t need a lot of GPUs)
GraphQL makes many REST problems disappear. Clients breaking after API changes, waterfall requests, over-fetching data, end point discovery, security and performance.
Being know via content / curation etc will not give you greatest ROI. Just spend that time learning and doing and get recognised with your work.
GetLago uses OLAP (online analytical processing) tools like Click-house to scale their events engine. They find OLAP tools are generally efficient at answering complex read queries and good at batch processing a large number of write queries
AI will accelerate the race between technology and education. To thrive, focus on generating real, hard-to-replicate value, stay curious and keep learning. Choose the right race to run. (Career advice for young high-performers)
Execution beats strategy every time. Good strategy is important, but execution is what matters. Focusing on execution also allows us to iterate toward the strategy that wins for customers today, even as the world changes.
Others
Meta used monolithic architecture to ship thread in five months
Intercom redesigned the company to be AI first
The way to get luckier is to find what’s repeatable (or simply do what works)