Rafael Cosman

2 years ago today, A16Z announced they put $100M into Pinecone, as everyone and their mother tripped over each other to invest in vector databases. Given how little standalone vector DBs are valued now, the hype cycle is wild to look back on. Quick timeline below: Late 2022: • ChatGPT launches. • Early RAG architectures emerge. • First buzz around Pinecone, Weaviate, Milvus, Chroma. Q1 2023: • Massive demand surge for vector search. • Vector DBs seen as critical infra. • Heavy VC excitement (“new database category!”). Q2 2023: • Peak hype. • Pinecone raises $100M, Weaviate raises $50M. • Every AI infra diagram has a vector DB layer. Q3 2023: • Saturation hits. • Postgres, Redis, MongoDB add native vector search. • Clouds bundle it. Q4 2023: • Hype fades. • Vector search becomes a commodity. Early 2024: • Market tightens. • Attention shifts to agents, fine-tuning, orchestration.

763

117 Comments

While OpenAI signed $1.15 trillion in compute contracts through 2035, DeepSeek trained a frontier model for $6 million. This was 2025’s central question : are we building on bedrock or quicksand? The top 10 posts of 2025 examined some of these topics : Are we in a bubble echoing the telecom crash, or building the next internet? Do traditional exit paths still work when secondaries dominate & IPOs vanish? How do you design tools when the user is AI, not human? 2025 forced a reckoning with reality. 1. How AI Tools Differ from Human Tools (https://lnkd.in/d9qcnbyz) : I consolidated my 100+ AI tools into unified, parameter-rich interfaces based on Anthropic’s research. The counterintuitive finding : AI systems need complex tools with complete context, while humans need simple, chunked interfaces. Claude’s success rate approached 100% after the redesign. 2. Back to Text (https://lnkd.in/dWcy53fv) : How AI Might Reverse Web Design : I watched an open-source agent book flights by navigating airline websites, extracting data from visual chaos. If AI thrives on pure text, the future of the web might look exactly like it started : simple text, but for robots instead of humans. The better AI performs, the fewer websites we’ll visit. 3. Circular Financing (https://lnkd.in/d8-KMGZf) : Does Nvidia’s $110B Bet Echo the Telecom Bubble? : Nvidia’s vendor financing totals $110B in direct investments plus $15B+ in GPU-backed debt, 2.8x larger relative to revenue than Lucent’s exposure in 2000. But unlike the telecom bubble, Nvidia’s top customers generated $451B in operating cash flow in 2024. The merry-go-round has paying riders. Read the full post here : https://lnkd.in/d8Xw6tJt

43

24 Comments

With RL techniques like GRPO, it’s become easy to fine-tune small models to outperform foundation models on vertical tasks. We’re open sourcing Osmosis-Apply-1.7B: a small model that merges code (similar to Cursor’s instant apply) better than foundation models. Links in comments! Using foundation models for high specificity, low complexity tasks like code merging can be overkill - instead, it’s better to use a specialized model instead for higher performance and latency. We fine-tuned the model on 100K real-world code commits where the model needs to merge a code edit snippet into an original code snippet. On the test dataset (subset of CommitPackFT), foundation model performance ranged between 0.77-0.93 reward score. In comparison, Osmosis-Apply-1.7B achieved a 0.98 reward score - while also being 3X-5X more affordable than the next cheaper model AND ~10X faster! Why this matters: the best AI agents use multiple models to ensure task specialization and reliability at scale - on top of cost and latency benefits as well. Links to the model and full write-up (with reward function, hyperparameters, etc.) are in the comments - reach out if you’re interested in using reinforcement learning!

239

18 Comments

GPUs have gotten faster, and models have gotten larger. But storage is still stuck in the past. Modern AI workloads need instant access to huge datasets. Instead, teams waste time on slow retrieval, idle GPUs, and costly pipelines. Archil is changing this. It delivers object-storage scalability at block-storage speeds. The platform allows for seamless, high-performance access to massive datasets. By eliminating cold starts and accelerating throughput 30X, Archil enables teams to fully utilize their infrastructure and move faster on training, analytics, and AI deployment. Archil’s founder Hunter Leath brings rare technical depth and operating experience from building Amazon EFS and optimizing Netflix’s cloud performance. This is a team building the foundational layer that AI workloads desperately need. I’m thrilled to lead Archil’s seed at Felicis with Nancy Wang and support their mission to reinvent cloud storage for the AI era. https://lnkd.in/gWueXhzz

65

2 Comments

Excited to announce Emergence Capital lead investment in Vesence's $9M seed round!   Legal AI should be a slam dunk. It’s a text-rich domain with tons of training data; well-suited for LLMs. Thus, it's no surprise that AI demand amongst lawyers has been extraordinary.   Unfortunately, in many scenarios, the reality hasn't yet delivered on the promise. Many AI tools fail for the same reason: they prioritize generation over precision.   Hallucinations and inconsistencies turn time-saving innovation into a liability and lawyers have to re-review every clause and citation, which adds hours of work instead of saving them.   Before AI, firms solved this with armies of reviewers scrolling through Word docs line by line.   Vesence is tailor made to bring precision to law at scale.   Instead of generating new text that might be wrong, Vesence uses AI to ensure what’s already written is right.   It enforces firm-wide standards, validates references, flags inconsistencies, and polishes language to perfection — all inside Word and Outlook.   Even more impressive than the product are the founders. Henrik and Ludvig embody the precision and persistence we look for at Emergence.   Henrik built the first version solo, including full Microsoft integrations. And Ludvig, a former pro athlete, brings the discipline of an elite competitor. Indeed, they literally slept at a top Swedish law firm to study the workflows firsthand as they built out the product. Intensity driving precision.   We're thrilled to welcome Henrik and Ludvig to the Emergence family.

328

16 Comments

RL-as-a-Service is emerging. Foundation models gave us a leap in generalization—but they’re static. Once trained, they don’t improve. That makes them poorly suited for workflows that require long-horizon reasoning, adaptation, and error recovery. Today’s agents still break easily. They can’t debug, replan, or persist. Prompting alone isn’t enough to teach real-world behaviors. What’s missing is memory, feedback, and the ability to learn from interaction—capabilities reinforcement learning provides. But RL today looks like the pre-GPT-3 era: narrow tasks, brittle fine-tuning, no scale. That’s changing. Dozens of startups are emerging the past few weeks to serve internal teams at the model providers by building the new substrate: simulated work environments, autograded evals, and scalable, economically meaningful task distributions. Think: training gyms for agents. Labs are paying close attention. Just as pretraining required billions of tokens, generalist agents will need millions of interaction traces across the workflows that dominate labor spend. This includes everything from email triage and customer support to form-filling, ticketing, and enterprise coordination. For startups, the implication is clear: brittle, hardcoded flows won’t hold. As generalist models trained through RL-as-a-Service improve, the next defensible layer will be reward design, eval engineering, and task simulation. The model matters—but the environment is what makes it useful.

355

26 Comments

I’ve been working on the theory of why small, shippable factories are the perfect inversion to the centralized prefab factory model. It all collapsed into one line of math (credit to Gilles Retsin): Factory Efficiency = Usage × (Throughput × Product Value ) / ( CapEx + Deployment OpEx) The idea is that the closer you can drive Usage to 100 % while keeping the denominator tiny, the faster the flywheel spins.  A shippable microfactory tackles this by primarily focusing on a high Usage factor. Because the microfactory can be packed up afterwards and reused, its CapEx gets amortized over multiple projects, effectively raising its overall Usage across the year. If one project alone doesn’t fully occupy the factory, it can simply roll over to another job to maintain high Usage. Its CapEx is often an order of magnitude lower than a centralized factory (< $1M vs. $10-50M min), so the breakeven throughput is more achievable on a small pipeline of work. The real breakthrough here is portability. There is near-zero stranded capital and almost no idle time. To flesh this theory out, I wrote 3000 words on the topic for Brad Hargreaves and Thesis Driven. The article breaks down: - why “fixed factories, shipped goods” is being inverted to shipped factories, fixed goods - what a sub‑$1 M robotic cell does to CapEx per home versus a £45 M off‑site plant - early production/cost data on a shippable microfactory from Mollie Claypool, Gilles Retsin, Sam Baker and the Automated Architecture (AUAR) team - the hybrid reality of robots tackling the heavy structural components and human crews assembling faster and cheaper - technical constraints to pull this model off and future considerations on the optimal business model (own vs. rent) Major shoutout to Gilles Retsin for helping me pull this together. https://lnkd.in/eRTnbZPa

54

2 Comments

🚨 Big milestone for one of TipTop VC’s earliest investments: GC AI just announced their Series A, led by Sound Ventures, with participation from an incredible group of operators and strategic investors. From the start, it was clear that Cecilia Ziniti and Bardia Pourvakil were building something different — not just another legal tech tool, but an entirely new way for in-house legal teams to get work done. In just over a year, they’ve gone from zero to: • 450+ legal teams across 53 countries • 600,000+ legal AI interactions • 3x revenue growth in 2025 alone We were fortunate to invest in GC AI at the pre-seed, and it’s been a privilege to watch this team execute with speed, focus, and deep customer empathy. This is why we built TipTop Ventures: → To partner early with exceptional founders reinventing work in large, complex industries → To help back the future of vertical software + AI, before it’s obvious → To earn a seat at the table with companies like this Congrats again to Cecilia and team — excited for what’s ahead. 🚀

174

25 Comments

Every Sunday, I recap the biggest themes in AI, tech, and markets. This week: Karpathy set the record straight, the AWS outage revealed system fragility, OpenAI’s Atlas challenged Chrome, Google claimed quantum advantage, and vibe coding met reality. ▪️Karpathy reframes progress as a grind that will take years of plumbing - every extra “9” of agent reliability costs blood, sweat, and server logs. ▪️AWS outage shows we’ve built a digital civilization on YAML and hope - when infra fails, so does modern life. ▪️OpenAI’s Atlas eyes Google Chrome’s throne - because in the age of agents, owning the browser means owning the interface of intent. ▪️With 13,000x speedups and verifiable outputs, Google is on the path to turn quantum computing from hype to hardware. ▪️Vibe coding hits its limit where syntax meets system design - turns out, you can’t prompt your way past architecture. 📩 Full post here: https://lnkd.in/gaZGtHNS

67

2 Comments

Despite all the big funding rounds and flashy demos in US robotics, K-Scale’s inability to raise more money should worry us. We're at risk of replaying the LLM story all over again in robotics: - Chinese companies are going open-source and collaborating across the value chain (from EV suppliers to downstream integrators) - most western teams are going full-stack proprietary, closed-source, all in-house Guess which robots the next wave of research labs and startups will actually be able to build on when they want to invent new algorithms or tackle unseen real-world use cases?

273

33 Comments