35,848 results · Most relevant
When you need to maximize revenue from a first deal (e.g., for a demo day), throw out a number in a way that does not get you slapped. The specific approach is to use a case study: present a semi-real example of what others have paid and let the buyer react, rather than the seller proposing pricing directly.
Work backwards from the case study you want to write about each prospect company — how you got them, how you sold them, how you delivered — to keep the sales motion consistent. Use the case study for qualification, disqualification, and helping people understand what you do.
Frame the early offering so customers are not expecting a full enterprise-grade platform ready next week; they expect something more service-shaped. Deliver it with a mix of manual and automated work to figure out what the actual thing to build is. Selling this service-shaped offering validates that there is real pull.
When the buyer gives a deadline (e.g., Q1 next year), treat it as a forcing function and work backwards from it to identify the steps needed to be deployed and showing results by then. That deadline is a forcing function for all the prop steps in the sales process.
Sofia Jakovcevic recommends using the prompt playground to test a single agent (a single prompt with a set of tools attached, a model, and LM configurations), and the agent playground to evaluate multi-agent systems, including very complicated multi-agentic systems. You can connect a remote agent via HTTP call so you can change how different sub-agents communicate with each other directly in your codebase.
Sofia Jakovcevic notes you can run experiments programmatically with the SDK, REST, CI, or CLI in addition to the UI. Her favorite method is with coding agents using agent skills, which allow you to connect coding agents with external databases and advanced context.
Sofia Jakovcevic demonstrates adding scoring mechanisms (evals) like task completion and action safety to playground experiments, and notes you always have the ability to customize your own evaluators in addition to pre-built ones like hallucination, user frustration, and security checks.
A guest recommends optimizing inference serving subject to the SLA: how fast the response needs to start and how quickly the whole generation must finish. They also recommend routing requests not only to balance load but also to the GPU where the KV cache is already resident, warning that round-robin load balancing alone is insufficient for LLM serving because it ignores KV cache residency. They further recommend serving pre-fill and decode as separate workloads, disaggregating them onto different GPUs with different parallelism or kernels, and ensuring correctness of tool-call tokenization, since tool calls introduce significant complexity at the tokenization stage.
Francois lists several ways to cut test-time flops: use MoE with K experts to reduce flops by roughly 1/K over all experts; use SSMs so full attention is only needed on a fraction of layers (e.g., every fifth or tenth layer); use multi-headed latent attention to compress in the sequence domain; use quantization, even down to 2-byte models; and distill a large model into a smaller model when the smaller one suffices for serving.
None stated; presented as endorsed engineering options.
Pull tool calling up into the harness and execute it consistently across providers to gain accuracy instead of delegating it to providers. Same with reasoning, extraction, and thinking.
None stated; presented as endorsed advice from Jeff.
Francois recommends several techniques to cut inference cost: keep cached pre-fill results hot in DRAM, Redis, or SSD and use an LRU cache for big savings; use SSMs to skip full attention on most layers (e.g., only every 5th or 10th layer) to cut test-time inference cost; use quantization, even down to 2-byte models; use radix attention (prefix/radix trees) for KV cache reuse, borrowed from classic OS and database theory; use page attention (virtual memory paging / buffer pool into DRAM), which is mutually exclusive from radix attention; and route requests so not everything goes to the big model, waterfalling or tool-calling up to bigger models to save flops.
Hari Kumar describes Contour's goal of building in-house: bringing all the tools different analysts use — like AlphaSense — into one internal platform that learns from analysts' habits, while up-leveling current spend; previously analysts just used the AlphaSense portal within their existing models.
Hari Kumar says Contour factors into its ROI equation the pauses from breakdowns (model calls or connectors failing every couple of days) and the time taken away from valuable insight work to fix bugs that the tech side cannot fix itself — noting this detracted time decreases over time but remains a tracked metric.
Hari Kumar shares a helpful practice: have an agent run expert call interviews, synthesize the notes yourself, and compare that against having the agent synthesize the notes before coming to your conclusions.
Dane Knecht advises prioritizing patching existing software first, describing the current phase as the industry 'patching 70 years of software' — catch-up before getting ahead of the new delivery pace. He also says not to freeze feature shipping: keep shipping new things while patching in parallel, because if you're not shipping, you're dying. AI can make both shipping and patching faster.
creator advice, endorsed
Document your work publicly — do a thousand outreach attempts or knock on 10,000 doors and film the process so people can watch you do work they don't want to do.
Use a trackable link-in-bio (e.g., /ig) per platform to measure conversion rates from each channel. Track the conversion rate of viewers who click your platform-specific link and download the app to identify your best channel.
Commercial context: The creator presents this as their own offering and directs viewers to get it.
Make videos optimized for specific platforms rather than using identical content everywhere. Twitter generally doesn't like vertical videos — it likes 16:9 — and Twitter videos should communicate in the first seconds without sound, using quick visual beats. Assume viewers won't read on-screen text; communicate through rapid visual demonstration instead.
When making a product demo video, show the product working in real time with no cuts to build authenticity. Counter to the assumption that Instagram requires quick edits, a zero-cut demo video got 4.6M views there.
Launch demos almost every single day rather than saving everything for polished launches. Do a polished launch only about once a quarter for something you're really proud of and want talked about properly.
Choose sources → to get a library of your own and a daily brief.