Report · MoneyThe other strategy in the market, which is what the labs do, is an FDE (forward-deployed engineer) approach: a human goes into your business, understands how it runs, then builds custom software with AI that uses the lab's AI. Enterprises are described as willing to spend around $5 million a year paying OpenAI for forward-deployed engineers.
the $5M figure is approximate ('like 5 million') and speaker-reported
The a16z Show · What Happens When Al Agents Can Spend Your Money?
Creator perspective · MoneyGaryVee shares that using AI tools like Claude, ChatGPT, and Gemini for research enabled him to learn in one year what would have taken 800 hours in only eight hours.
GaryVee · The Biggest Opportunity Most People Are Ignoring
Creator perspective · MoneyA decision model can check whether a real-world state passes a contract defining the ideal state; a natural fit is on the back end of the hosted run.prose as a verifier at the end of the job that is very fast and cheap.
AI Engineer · Stop Telling Agents How — Dan Barrett, Raymond Weitekamp & Jose Montes de Oca, OpenProse
Creator perspective · MoneyJev becomes a very cheap semantic classifier or semantic memoizer — a cache that decides when something is out of date and kicks off everything downstream; the event-driven half works like React-style memoization, and an earlier agent-based state-change detection system was very expensive to run.
AI Engineer · Stop Telling Agents How — Dan Barrett, Raymond Weitekamp & Jose Montes de Oca, OpenProse
Report · MoneyEvan Conrad describes SF Compute's programmable SKU concept: customers use APIs to define what they want (e.g., willing to buy Blackwells at $6/hour or less), and the order book matches that query against any composition of sellers that can fulfill it, such as GB300s or B300s; named queries are a new feature but the query system is how the order book already works.
Latent.Space · The Programmable Data Center — Evan Conrad, SF Compute
Report · MoneyEvan Conrad describes SF Compute's offering: customers can buy compute for any window from the next hour up to three years out; the order book contains blocks of capacity extendable to any arbitrary time horizon; currently blocks are paid entirely upfront, unlike longer-term leases.
Latent.Space · The Programmable Data Center — Evan Conrad, SF Compute
Principle · MoneyIn a measured example with B200 GPUs serving the same model (~8,000 input tokens, ~1,000 output tokens, P50 target of 100 output tokens/sec per user), one serving stack delivered ~200 generated tokens/sec at P90 while another delivered ~300 — the same chip at the same target interactivity can deliver about 50% more output depending on the operator. The difference comes down to software, settings, and number of requests running — all operator choices.
AI Engineer · Same GPU, 50% More Output: Why the Operator Matters — Oliver Kennon, SemiAnalysis
Report · MoneyMonzo employees can invoke Agent Chip by reacting to any Slack message with a specific emoji, which the agent picks up as a prompt. Added to Slack, it answers questions, explains architecture, and chats. It performs code reviews like a very senior engineer ('it looks good to me') and can be assigned Linear tickets, which it works on autonomously.
AI Engineer · A Bank Built Its Own Coding Agent. It Took Off — Felipe Felisola Caso & Francesco Galletta, Monzo
Principle · MoneyCard rewards are funded by a cross-subsidy: issuers offer a big headline reward on one category (like dining) while the cardholder also spends on other categories (like fuel). When agents optimize for points across all transactions, that cross-subsidy model collapses and the whole points reward ecosystem falls apart.
described
The a16z Show · When AI Agents Start Spending Your Money | Alex Rampell & Terry Angelos
Creator perspective · MoneyThere are now two distinct kinds of SEO: SEO targeting human customers and SEO targeting AI agents to get the agent to pick your solution over competitors.
Startups For the Rest of Us · Episode 855 | The Eight SaaS Moats that Still Work in the Age of AI
Report · MoneyIn February 2025, public markets dropped due to lack of confidence in SaaS and cloud companies, an event the hosts call the SaaSpocalypse, and the shock rippled through M&A markets. The episode was recorded in October, several months later.
Startups For the Rest of Us · Episode 855 | The Eight SaaS Moats that Still Work in the Age of AI
Creator perspectiveWhen dealing with a new benchmark or problem of economic value, the guests say they have not found a replacement for what they call quality time: literally running hundreds of experiments, looking at them, developing an intuition, and trying to figure out what is wrong and what is not.
presented as the guests' practice
The a16z Show · The Race to Recursive Self-Improving AI
PrincipleOwen explains that when proving a recursive self-improvement system works, you should pick a problem domain as close as possible to AI itself so improvements feed back into the system.
The a16z Show · The Race to Recursive Self-Improving AI
Creator perspectiveThe guests describe the nano-GPT benchmark of retraining GPT-2: originally a very expensive model taking weeks on multiple GPUs, it can now be trained on a single GPU within hours for roughly a hundred dollars. They emphasize the gains come from the same GPU and same data — improved pre-training techniques, algorithmic techniques, data techniques, and kernels — and note that many of these improvements are auto research results produced by agents rather than humans.
speaker hedged on the weeks figure; cost figure corrected mid-conversation from 'a few hundred dollars' to 'a hundred dollars'
The a16z Show · The Race to Recursive Self-Improving AI
Report · MoneyAnalyzing his own agent sessions from 2026 year to date, Corey Gallon found that in about 4 in 10 sessions the agent got something wrong that mattered.
AI Engineer · I Believe Everything My Agents Tell Me: Three Levels of Review — Corey Gallon, Rexmore
Report · MoneyThe three review levels have been packaged as agent skills and released as open source. The write-up is at gallon.me; pointing your agent at the write-up will let it know how to check its own work using this review method.
AI Engineer · I Believe Everything My Agents Tell Me: Three Levels of Review — Corey Gallon, Rexmore
ReportAI systems have been deployed as safety nets to clinicians, flagging when something might be off in the moment; in the PandaHealth study, clinicians prone to antibiotic overuse were flagged, encouraging them to think about the global issue of antibiotic resistance rather than just the local case.
described
The a16z Show · What Happens When Everyone Has a Health Expert in Their Pocket? | OpenAI’s Karan Singhal
Creator perspectiveOpenAI is training medical expertise into its general purpose models rather than building a separate healthcare model — the same model people use to write code or plan a vacation. Karan Singhal, who previously worked on specialized models and has experience on both sides of the question, says OpenAI's approach is specialized training on a general model, which he considers the best of both worlds. For all models shipped in ChatGPT in the last year, specialized healthcare data has been added to each phase of model training. The result is a model that benefits from the scale and general improvements in intelligence and reasoning of big frontier models plus the benefits of specialized training from integrating data from working with physicians across the globe.
described approach
The a16z Show · What Happens When Everyone Has a Health Expert in Their Pocket? | OpenAI’s Karan Singhal
Creator perspectiveIf AI pre-processes patient cases and patients pre-research, visits can focus on the human aspects of healthcare — empathy, bedside manner, and physical interaction — though visits still exist to gather information through history taking and investigation, whose physical elements will not go away.
hypothetical
The a16z Show · What Happens When Everyone Has a Health Expert in Their Pocket? | OpenAI’s Karan Singhal
Creator perspective · MoneyNick describes setting up a report to catch anyone who came from Gibson Dunn — a customer of his — and moved to Jones Day, an account he was trying to break into. This operationalizes the past-user tactic.
30 Minutes to President's Club | No-Nonsense Sales · #618: LinkedIn Sales Navigator: How I Get Warm Intros Every Week