Recommendation · MoneyFrancois advises keeping cached pre-fill results hot in DRAM, Redis, or SSD and using an LRU cache for prompt reuse. He also recommends using radix attention (prefix/radix trees) for KV cache reuse, drawing on classic OS and database techniques.
None stated; presented as endorsed engineering advice.
Also discussed by AI Engineer (unclassified source — not counted until classified). Not counted as independent sources.
Y Combinator · Not All Open Model Tokens Are the Same | YC Paper Club
Recommendation · MoneyUse logprobs to measure drift from a provider by comparing next-token probability distributions against a previous baseline. Send a single-token prompt and measure the log props coming out the other side against a baseline to test inference quality over time. Test next-token accuracy with progressively longer prompts to measure deviation in log probs as context grows. Note: logprobs are not available on many APIs, only detect drift from the past (not absolute correctness), and aren't helpful if you're starting from a place that wasn't serving the model correctly.
Limitations explicitly stated: API availability, drift-only detection, unhelpful if baseline was wrong.
Y Combinator · Not All Open Model Tokens Are the Same | YC Paper Club
Recommendation · MoneyWhen optimizing inference for coding workloads, analyze the cache hit rate and the turn latency (time between user turns) to determine responsiveness requirements. Also measure how long the user waits between turns and how fast generated code must complete, subject to an SLA, before optimizing the serving stack.
None stated; presented as endorsed advice.
Y Combinator · Not All Open Model Tokens Are the Same | YC Paper Club
Recommendation · MoneyJeff recommends using logprobs to monitor provider drift by comparing the next-token probability distribution against a previous baseline.
Y Combinator · Not All Open Model Tokens Are the Same | YC Paper Club
Recommendation · MoneyUse a distribution test comparing outputs of a reference model and a provider API to measure the gap between their output distributions. A distribution test doesn't catch intricate problems like poor quantization and doesn't tell you why something is going wrong, unlike a comprehensive eval suite where you can pick out a specific test. It requires the reference model running, which is an extra expense.
Limitations explicitly stated: misses intricate problems, no root cause, requires reference model.
Y Combinator · Not All Open Model Tokens Are the Same | YC Paper Club
Recommendation · MoneyJeff recommends measuring the deviation in logprobs as the prompt gets longer, to see how next-token accuracy degrades with context length, and testing that next-token accuracy past the 200,000-token range remains consistent with the reference model provider's output.
Y Combinator · Not All Open Model Tokens Are the Same | YC Paper Club
Recommendation · MoneyJeff recommends a distribution test that compares the outputs of a reference model and a provider API to measure the gap between their output distributions as a cheap provider-quality check.
Y Combinator · Not All Open Model Tokens Are the Same | YC Paper Club
Recommendation · MoneyBecause pre-fill and decode are very different workloads, serve them in different ways: separate (disaggregate) them onto different GPUs and consider different parallelism or kernels for each. Disaggregation choices affect user experience and total GPU throughput.
None stated; presented as endorsed advice.
Also discussed by AI Engineer (unclassified source — not counted until classified). Not counted as independent sources.
Y Combinator · Not All Open Model Tokens Are the Same | YC Paper Club
Recommendation · SocialCoach Kyle notes logistics matter (the woman was leaving in a day and a half, visiting, heading to a cruise) and that given poor logistics and lukewarm interest he transitioned to 'let's just keep in touch.' He adds that if he had wanted to pursue her, the play would be a longer, flirtier 5-10 minute interaction for a solid number — and that rejection is no big deal: accept it, wish her well, and move on.
Coach Kyle · A Realistic Approach on the Train (Infield)
Recommendation · SocialCoach Kyle recommends smiling during the approach and using a genuine (not nervous) laugh out loud when a woman's family is present, as it disarms the family and the woman.
Coach Kyle · A Realistic Approach on the Train (Infield)
Recommendation · MoneyHarneet Kaur recommends taking all the data for a chosen company and running it through the smartest AI models as an interesting exercise. She chose iHeart Media as the test company — a company that filed for bankruptcy in 2018 and did another stress transaction in 2024 — and ran the situation through Astra 6 on AXM Intelligence and Opus 5.5.
AI Engineer · Can AI Understand a Distressed Credit? - Harneet Kaur, Axim Intelligence
Recommendation · MoneyDane Knecht recommends letting employees choose the AI models and tools they see fit rather than mandating a standard stack, and says it's silly to ask humans to decide which model to use for each task — tasks should be intelligently routed (frontier models for planning, cheaper open-source models for execution, frontier models again for review and red teaming). Cloudflare doesn't grade employees on AI token usage; they're graded on what they've shipped and built.
creator advice, endorsed with high confidence
No Priors: Artificial Intelligence | Technology | Startups · Democratizing Security for the Agentic Web with Cloudflare CTO Dane Knecht
Recommendation · MoneyDane Knecht recommends scaling security teams by going on offense — expanding from one red team to four or five that actively try to find vulnerabilities — and focusing security effort on the discovery side rather than perfecting patching processes, since engineers will automate patching at scale over time.
creator advice; red-team count is the speaker's expectation
No Priors: Artificial Intelligence | Technology | Startups · Democratizing Security for the Agentic Web with Cloudflare CTO Dane Knecht
Recommendation · MoneyDane Knecht recommends that content site owners know their CPM (revenue per thousand ad impressions) and set an agent-access price based on it, so an agent can decide whether to read the content and pay that CPM rate. He envisions a world where agents can make that decision and pay directly.
hypothetical future scenario presented by the speaker
No Priors: Artificial Intelligence | Technology | Startups · Democratizing Security for the Agentic Web with Cloudflare CTO Dane Knecht
Recommendation · MoneyAndrey Rybka suggests checking whether your agent can access DNS as one of the security homework questions.
AI Engineer · Real AI Agent Attacks and What Would Have Stopped Them — Andrey Rybka, Bloomberg
Recommendation · MoneyWhen negotiating GPU deals quickly, clarify what 'memory' means — VRAM, CPU memory, or additional storage — because it affects what loads can be run, end-user needs, pricing, and delivery dates.
AI Engineer · Why Compute Still Trades in Group Chats — Affaan Mustafa & Alejandro Soumah Fernandez, Jump Trading
Recommendation · MoneyAfter making your first couple of videos, start asking analytical questions about conversion per channel: use a trackable link-in-bio (e.g., slash IG for Instagram) to measure who downloads the app, and offer a button that lets viewers send the app to themselves over email so you can track clicks and conversion.
Commercial context: The creator presents this as their own offering and directs viewers to get it.
Y Combinator · This Is How You Should Actually Make Launch Videos
Recommendation · MoneyMake videos for specific platforms — most are made for Instagram then posted to Twitter, some the reverse. Twitter generally doesn't like vertical videos (it prefers 16:9), and videos should be optimized for muted, fast-scrolling viewing with quick visual beats so nobody has to read the entire thing.
Y Combinator · This Is How You Should Actually Make Launch Videos
Recommendation · MoneyOpen the video with a hook like 'Hey, no hands' — showing something unexpected right off the bat gets people to watch. Structure demo videos as a story with escalating beats: play the song, lower the music so you can keep talking, then pull up the Stripe dashboard so viewers wonder how much money was made.
Y Combinator · This Is How You Should Actually Make Launch Videos
Recommendation · MoneyWriting is so important — spend as much time as possible before the video writing it. But if you've never made a video, don't worry about writing before recording; if you're on your 15th, 20th, or 30th video, try writing for just 10 minutes about what the video should be and then record it. The more you think through the video before you get to the camera, the better it can possibly get, because you can catch a bad idea before recording. For his writing he uses FreeWrite, a free open-source app at freewrite.io that he built himself.
Commercial context: The creator presents this as their own offering and directs viewers to get it.
Y Combinator · This Is How You Should Actually Make Launch Videos