Jason Gilbertson

AI has entered its systems phase. Outcomes are no longer driven by model size alone, but by how well models are integrated into context, tools, routing, and feedback loops. A small set of primitives is defining this shift: - Tool use turns models into operators, not just assistants. - Dynamic context (e.g. MCPs) makes context portable, composable, and governed at runtime. - Intelligent routing allocates the right model to the right task under real-world constraints. - Telemetry closes the loop, turning production behavior into learning. On top of these primitives, new alignment strategies – synthetic data, verifiable rewards, and multi-turn objectives – are allowing systems to learn from real outcomes rather than static labels. The implication: As capabilities commoditize, differentiation moves up the stack to system design, domain grounding, and alignment with measurable value. Pricing follows the same path, shifting from labor proxies to outcomes, and in some cases even risk-sharing. For builders: think more about control planes that govern routing, context, and accountability. For buyers: evaluate AI as an operator, one that can be audited, improved, and ultimately trusted with responsibility 🔗 More here: https://lnkd.in/gYfF7cWu

23

1 Comment

Iliya Valchanov's latest piece nails it: "A good prompt engineer is a good domain expert." ... because they know the context. They know when AI output is subtly wrong...or very wrong. But here's what I'm seeing with startups we work with: nobody wants to hire juniors anymore. Everyone wants someone ready-made—knows AI, immediately productive, a whizz at everything. The problem? Knowing AI doesn't make you a domain expert. If we stop hiring juniors because AI handles entry-level work... and juniors never develop into domain experts... who's left to direct the AI in 10 years? We might be undermining the very thing that makes AI useful. The expertise that powers Human+AI takes years of repetition that happens in junior roles. Roles we're eliminating. Or maybe expertise gets built differently now. Maybe AI itself becomes the training ground of its own domain expertise and at some point it will not need anyone to prompt it. A true Agent from inception and manifestation, not just through execution. I don't know. But it's worth sitting with. What's your read? (and P.S. Do read Iliya's blog - totally worth it + great hands on materials inside). https://lnkd.in/dSSXJupm

24

4 Comments

What if you could search complex documents with charts and diagrams as easily as you search text? This Vision RAG tutorial shows how, no OCR, no complex parsing, just multimodal embeddings doing the heavy lifting. Full tutorial: https://lnkd.in/eT9u_J7q

10

The more you round out coverage within a semantic space, the more you show up. It's that simple (and that hard). Think about it like this: when the model thinks about burgers, it probably goes McDonald's, McDonald's, McDonald's, McDonald's. Oh, we should recommend McDonald's. The goal is to be that present in the model's understanding of an industry and what a company does. New content helps with this. But here's the catch: it needs to be built with current best practices for those search result bots. When they land on a page, give them clear signals: "If you stay here, this is what you're going to get." Then they can decide if the content is worth processing and potentially citing. And if you're going to give them information, make it: 👉 Bite-sized and chunked (scannable) 👉 Backed by evidence 👉 Specific (not vague claims) If it's not specific or credible, the bot will skip it and find somebody who is. So yes, create new content to fill gaps in semantic coverage. But do it strategically, with AI readability in mind from the start.

7

3 Comments

I expect 2026 to be the year we start to see some prominent enterprise AI IPOs, notably OpenAI, Anthropic & Databricks. Tasmin Lockwood of CNBC Europe asked me my thought on the first two of those as well as SpaceX. I also commented on the use of the proceeds of any IPO: '“We expect a significant portion to go toward custom silicon development and energy infrastructure. The IPO that outlines the clearest path to owning their own compute stack will be the winner in the long run,” Patience said, noting that this reduces a company’s reliance on hyperscalers.' More here: https://lnkd.in/egPTDqZR

17

1 Comment

Three headlines from this week tell one story: → Andrej Karpathy's "autoresearch" runs 100 ML experiments overnight while you sleep — 630 lines of code, one GPU, fully autonomous → Anthropic's new Code Review deploys a team of AI agents on every pull request — PR review coverage jumped from 16% → 54% internally → LangChain's Harrison Chase: better models aren't the bottleneck. Building the right "harness" around them is. The through-line? AI is no longer just completing tasks. It's owning loops. Research loops. Code loops. Review loops. The human role is shifting: from doing the work, to designing the system that does the work. Karpathy put it bluntly — humans are slow "meat computers" running on sleep and group meetings. Agents don't have that problem. The competitive edge isn't going to "use more AI." It's going to the teams who #design feedback loops clear enough for an agent to optimize inside of them. I'm at the Gartner Data & Analytics Summit in Orlando this week — and this is the conversation happening everywhere. If you're here and want to grab a coffee today, send me a DM. ☕ Thanks Erick Brethenoux for your Future of AI preso yesterday. #AIAgents #GartnerDA #DataAnalytics #GenAI #EnterpriseAI #ClaudeCode #LangChain #ArtificialIntelligence (links in comments)

14

3 Comments

How do you track sentiment with a machine that is not sentient? sentiment in LLMs seems largely a function of the input prompt and the personality programmed into the LLM's response style. Genie Jones has gone down this rabbit hole. We quite regularly get asked about sentiment tracking in AI output. Before tracking sentiment, some serious thought needs to go into what you want a tool to build, and what you want to do with the information. We'll get on to that. But in the meantime, this is why rushing in and "tracking something" might not be the best policy. https://lnkd.in/eM2CZYth

11

2 Comments

Wow, OpenAI dropped open-weight models (gpt-oss-120b and 20b), and honestly, this isn’t your average “hey we open-sourced something” hype. The big one (120B) is right up there with the best—saw MMLU at 80.1%, pretty nuts AIME math scores (up to 98%), and strong on those science benchmarks everyone talks about. Even the smaller 20B is surprisingly capable, and apparently you can run it on a machine with just 16GB RAM—thanks to lesser active weights. What’s cool: these are actually useful for real projects. Commercial-friendly license, full weights, and you can deploy them wherever—cloud, local, whatever. Not the usual roadblocks or gotchas. Been talking to my tech friends and the buzz is real. Feels like we finally have legit open models you’d actually choose instead of just tinkering out of curiosity. I ran the 20B on my mac and already love it, my genuine reaction is “okay, this changes things.” People are already talking about plugging these into regulated industries like Finance, startups, even healthcare—pretty wild. Bottom line for me: feels like a game-changer. Local, truly open, with actual quality you can trust—and not just for hobbyists. Super curious to see how fast these get adopted and what wild use cases pop up in the next few months. Quick nerd stats if you care: 120B: 80.1% MMLU, 98.7% AIME, 80.1% GPQA 20B: 71.5% GPQA, 90–98% AIME, and generally blows away old models this size Anyway, if you’re building with AI, definitely worth a look - https://lnkd.in/dNehh_nz Wild moment.

40

A video is circulating on social media of a robot running into an oversized mirror. Want to know why the robot did that? It likely wasn't trained to recognize that the mirror was a solid piece of reflective glass. The robot probably saw the frame around the mirror simply as a doorway into another room. Which is why, while physical AI companies are racing to collect more data, we're working towards meaningful data.

3

Data from our benchmark on 80,000+ human and AI review comments! What's really interesting to me is that GPT 5.2 and Gemini 3 Pro, with minimal prompting, produce extremely structured review comments, nearly as structured as R3 despite R3 being multi-agent. We defined the structural elements of a peer review comment: - the specific issue identified (specificity) - why it is a problem (rationale) - how to fix it (actionability) - where it is in the text (anchoring) We pulled these out of every human and AI review comment and measured their rate of occurrence per paper.

18

Very proud to announce Buoyant Ventures latest investment in Hammerhead to help unlock more power for AI and data center growth. My partner Laura Katzman (Dyer) led this deal and our research into efficient computing which helped us win this competitive and over-subscribed deal. Thanks to Maeve Allsup for your excellent explanation of the opportunity. https://lnkd.in/gSM4PadB

64

2 Comments

Sean M. Kerner: "[Pinecone] today announced Nexus, which it positions as a knowledge engine rather than an improvement on retrieval. Nexus introduces a context compiler that converts raw enterprise data into persistent, task-specific knowledge artifacts before agents query them, and a composable retriever that serves those artifacts with field-level citations and deterministic conflict resolution. "Alongside Nexus, Pinecone is releasing KnowQL, a declarative query language that gives agents a vocabulary to specify output shape, confidence requirements, and latency budgets. In Pinecone's own internal benchmark, one financial analysis task that previously consumed 2.8 million tokens was completed by Nexus with just 4,000. This represents a 98% reduction, although the company has not yet validated it in customer production deployments. Nexus is in early access starting today." https://lnkd.in/gHne_DGZ

17

1 Comment

Big shoutout to the OpenTSLM team for releasing their new paper on Time-Series Language Models. Exciting to see time treated as a first-class signal – a huge step forward for anyone working with dynamic, operational data. At Harmonyze we’re excited to see approaches like this gaining traction. The ability to reason over evolving patterns, surface inflection points, and generate adaptive recommendations opens up a ton of potential for real-world decisioning. Big moment for the ecosystem! https://lnkd.in/eh8XfFEc

27

🔥 🔥 The next wave of vertical AI winners won't be systems of record 🔥 🔥 They'll be systems of love. This is the blog post I haven't written, but reveal in this fantastic episode of Verticals w/ Luke Sophinos and Nic Poulos. I've been investing in vertical software for +15 years — back when my partners looked at construction project management software out of Santa Barbara and I was the only one in the room saying "no, this is legit." (That one became Procore.) Here's what I've learned about which companies actually win in markets with legacy incumbents: Law firms despise their Westlaw and LexisNexis subscriptions. Insurance agents hate Vertafore. No banks says I'm glad I'm built on Fiserv. They've had no alternative — so these vendors just doubled the inflation rate every year and called it a price increase. Now companies like Harvey goes in and asks: "What problems do you have?" The law firms say: "Where do we start?" and provide the keys to all their data and systems. The incumbents were systems of record. The challengers are building systems of love. When customers are that pent up, it's not really a product competition. It's a permission slip that you need to succeed in the AI world, because access to systems and data is the point of differentiation. The building of software has become the commodity. Link to the episode in the comments.

67

19 Comments

There's no shortage of flashy social media posts making AI transformation look easy. Stories of every employee shipping production-grade SaaS, overnight culture change, and frictionless adoption. From where I sit, that's not the reality. Getting people across teams, technical backgrounds, and experience levels to build meaningfully with AI is genuinely hard. The Hackathon we ran at Alpine last month is the most practical thing I've seen that actually moves the needle. This was a 3-week full firm effort that culminated with a competitive March-Madness -style demo day at our annual retreat. A few things we did differently that I think are worth sharing: — We started by crowdsourcing ideas by function across the firm (investing, sourcing, portfolio operations, internal operations, etc.) and had people vote on the ideas. Letting people loose without a filter tends to produce narrow, low-impact tools. The voting process helped surface the highest impact builds with real firm heat — We assigned cross-functional teams deliberately. Alpiners working on the same lived problem from different angles — We didn't just hand people Claude accounts and wish them luck. Before and throughout the hackathon, we hosted firm-wide AI upskill sessions with Alpine team members walking through hands-on, tailored trainings on how to actually build on our stack — Unlike a traditional hackathon designed for developers pulling all-nighters, we designed this for the new kind of builder: non-technical operators and investors who prompt LLMs, wait for agents to run, review the output, gather feedback, and iterate together. So instead of cramming everything into 1-2 days, we gave teams multiple weeks to get it right prompt-by-prompt — Then we reserved the hackathon itself for what actually moves people: live demos in front of the whole firm. The energy and inspiration people took from each other was a huge highlight — We set the bar high going in: come ready to ship. Not a prototype or a proof of concept. A real tool someone at Alpine could use on Monday. Not every build crossed that line - and that's okay! The high bar forced teams to downscope to something real and executable rather than overreach and deliver nothing What we ended up with: 30+ real tools headed for distribution across the firm, and 150+ people who now know they can build and are fired up for more. Huge thank you to Leeraz, Jack, and Adrianna for helping make this real, and to every Alpine builder who rolled their sleeves up. More to come!

47

2 Comments

If you are actually interested in AI, ML (the technology that must not be named - chuckle) and other application specific hardware performance numbers, I highly recommend checking out Puget Systems blog posts. Why? Well, in an age where everyone is shouting, and therefore no one is listening the company do a remarkably good job of getting into the performance details of various hardware configurations. In short, they are worth listening to because the provide nuanced and detailed orientated prose. It is almost like they have thought about things and tested their assumptions. Chuckle. Case in point: https://lnkd.in/g6ak9m4Z PS. They build best in class application specific PC’s.

9

2 Comments

🚀 Stop treating AI as just a "faster autocomplete." Welcome to the era of **Agent-Native Engineering**. A new operating model is completely restructuring how software teams work. Instead of just helping you type, AI agents are now acting as actual contributors—taking on scoped tasks, opening pull requests, and iterating on feedback asynchronously. Here is what the new engineering workflow looks like: 🔹 **Simple tasks** ➡️ One-shotted by agents from a clear prompt. 🔹 **Manageable tasks** ➡️ Delegated to background agents with built-in review cycles. 🔹 **Complex tasks** ➡️ Kept with engineers using synchronous coding agents. The result? Engineers spend less time writing boilerplate code and more time doing what matters most: scoping problems, reviewing work, and shaping the product. 💡 Check out this great read from Cofounder on the future of software development: https://lnkd.in/dJX5phfz #SoftwareEngineering #ArtificialIntelligence #AIAgents #FutureOfWork #AgentNative #TechTrends #SoftwareDevelopment

4

In the latest Greylock Partners Change Agents I spoke with Arvind Jain from Glean and Ping Wu from Cresta, two companies that are leading some of the largest agentic AI enterprise deployments. A few themes we dug into: - Why agent performance depends less on model improvements and more on enterprise context (systems, permissions, tribal knowledge - see Arvind’s follow-up post about this topic) - How agent sprawl is emerging inside large companies (Glean has customers with up to 2,000 unique agents built) and the subsequent orchestration needs - How they are building an ‘enterprise brain’ by observing human workflows over time, connecting various systems of record, and embedding that context into every agent Both Glean and Cresta started years before “agentic AI” was a term, which means enterprise understanding is the strong base upon which their agents are built. If you’re thinking seriously about agents in enterprise environments, check out the full episode below.

177

5 Comments