David Koblas

I write a short piece like this every weekday (sign up to get it in your inbox at tbpn.substack.com): There are two types of reactions worth tracking when a new model releases these days. The folks who actually use the model and can evaluate or benchmark it, and then the other labs. Let’s start with the more direct reaction. Gemini 3 is a newer, better Gemini that’s smarter, faster, stronger, newer, and better. The model releases are starting to feel a little bit like iPhone releases in that way. Everything is definitely better, but we aren’t seeing hugely surprising “how did they pull that off” type of demos come out nearly as often as they did a few years ago. The benchmarks are impressive though. Gemini 3 hit double the current state of the art on the very difficult ARC v2 benchmark. There are a bunch of interesting findings as you dig into how Gemini works through particularly hard puzzles and it’s still not surpassing human level performance at this particular benchmark, but it’s impressive progress. In general, Google has been fantastic at delivering models right across the Pareto frontier for over a year now. At every benchmarkable level of performance, Google is going to offer a very competitive model. Their vertical integration across the DeepMind team, their TPU chips, and Google Cloud Platform continue to compound and deliver a very competitive foundation model product. Consumer adoption is a bit more complicated, since there are a variety of ways users get locked in. Little product features like smoother speech-to-text, easy navigation of previous chats, and maybe in the future better collaboration all add up to maintaining a user base even if a better experience exists elsewhere. Watching the indirect reaction, from the other labs, might actually tell you more about Gemini 3 than anything else. Anthropic took a page out of Sam Altman’s book and announced a big strategic partnership. Anthropic will spend $30B on Microsoft cloud compute. Anthropic gets a $10B investment from Nvidia and $5B from Microsoft. It’s the sort of circular deal that we’ve already hashed out a few times, but this deal seems to be missing a bunch of zeros and somehow that makes me more optimistic about it. OpenAI used to be really aggressive about front running Google PR moments, but it’s been a bit quiet. A new feature that allows you to spin up group chats in ChatGPT launched five days ago, and Fidji Simo did a profile in Wired which I took as another push toward advancing the ads product. Personally, I think ChatGPT is incredible at designing bespoke gift guides and will probably be responsible for influencing billions of dollars of gift buying this holiday season. Unfortunately, I don’t think their agentic commerce product will be anywhere mature enough to actually take a meaningful share of this. __________________________________________________ Continue reading today's piece in the comments below.

5

1 Comment

Spec-driven development has roughly doubled my coding output over the last couple months. It's also created a ton of documentation exhaust. Most discussion about spec-driven workflows ends at "ship the feature." Nobody talks about what to do with the spec after. I've iterated several times on this — from deleting specs immediately, to keeping a graveyard folder, to a new approach I'm pretty excited about: graduating spec content into permanent documentation when work ships. I wrote up a post covering what we learned through all three iterations, plus the templates and skills we're using in production today. Link in the first comment.

25

9 Comments

“Every engineer remembers their first time.” Not shipping code—but watching it break in production at 3 a.m. And diving into a Slack war room with 50 others, trying to find a root cause in a sea of telemetry noise. That’s the world Traversal is rebuilding—from the ground up. At this year’s RAISE Summit at the Louvre, Anish Agarwal, Traversal’s CEO (MIT PhD in causal ML), took the stage to challenge one of software’s ugliest truths: the more AI writes our code, the less context engineers have—and the harder debugging gets. But Traversal’s thesis is clear: → Writing code is easy. → Keeping software reliable is the next trillion‑dollar problem. → And AI agents + causal inference are the missing layer. Instead of dashboards and “alert storms,” Traversal uses swarms of LLM agents performing parallel, semantic investigations, grounded in causal ML, to pinpoint the real root cause—fast. The company launched with $48 million in funding, led by Sequoia and Kleiner Perkins, with participation from NFDG and Hanabi. DigitalOcean has already rolled it out org‑wide, with AmEx and Eventbrite next in line. Considering enterprise downtime can cost up to $1.9M per hour, this isn’t just a better observability tool—it’s the AI SRE every engineering team will need. If GitHub Copilot is AI for creation, Traversal is AI for continuity—and it might just be the most important devtool startup of this generation. https://lnkd.in/e2eDTDC9

18

1 Comment

CRV Security: Request for Startups I never know if this actually works for our friends over at YC but figured we'd try. Here's what we want to fund in 2026! 1. Golden Artifacts: Think Chainguard but more broad. Artifact attestation exists for open source. Almost nothing exists for internal software — especially the vibe-coded tooling now running in production. We want the company building cryptographic proof of secure software delivered from secure artifacts: who built it, how, and whether it was reviewed. If more things are being yeeted into the world via Claude Code (myself included), this feels like an issue. 2. MCP & Agentic Security: Agents are getting real credentials and taking real actions. The security posture of most orgs around this is basically zero. That changes fast. You'd never give an employee hardcoded API keys or write access to your email without supervision/trust. Why give it to agents? 3. AI Governance: Boards are asking CISOs to account for AI risk. CISOs have no good answer other than "Palo has a module" 4. Next-Gen Endpoint: CrowdStrike was built for a world of static binaries and human operators. AI workloads, cloud-native infra, and AI-assisted attackers need a new architecture. The category is ready to be reinvented. 5. Networking in the AI Era: Zero trust was designed for humans. What does network security look like when the entity requesting access is an agent? Nobody's really solved this. 6. Email Security + Next-Gen Phishing: LLMs have made spear phishing infinitely scalable. I've never truly understood why Abnormal and KnowBe4 aren't one company. Maybe this time it's different. 7. Frontier Security Lab: We'd back a credible, well-staffed lab focused entirely on red-teaming models and setting the evidentiary standard the industry needs as LLM built apps become the norm. 8. Dependency Security: That Actually Remediates Malicious and vulnerable dependencies are a top attack vector. The tooling is mostly noise — scanners that don't close the loop. The winner here ships fixes, not just alerts. 9. Critical Infrastructure Cyber: Data centers, satellites, power grids, undersea cables. The physical backbone of the internet is increasingly exposed and wildly under-defended. We have data centers in space, for God's sake. Surely we need better cyber for critical infrastructure? 10. PAM for the Modern Era Legacy: PAM was built for static roles, human users, on-prem directories. Cyberark was founded in 1999.....Agents, ephemeral workloads, and cloud-native infra have broken all of those assumptions. Is anyone rebuilding this from scratch? If you're building in any of these areas — or something we haven't thought of — reach out. james@crv.com

452

49 Comments

David George at a16z published a piece yesterday arguing that there are only two paths left for software companies: grow fast or cut to the bone. Grow revenue by 10+ points or hit 40%+ real operating margins. No middle lane. He's writing about public software companies. But if you're running a 15- or 50-person company that uses software — not one that sells it — the implications land even harder. Here's what I mean. The big software companies you depend on are heading down one of these two paths. Some will ship AI-native products that change how their tools work. Others will cut everything that isn't core and raise prices on what's left. Either way, the tools your business runs on are about to change significantly. And that's just the vendor side. The bigger question is what's happening inside your own company. George mentions that the best engineers are now managing 20-30 AI agents simultaneously, and that four-person teams are outperforming departments. These aren't predictions — this is what's already happening at companies with the resources to move fast. For a smaller company, the math is more interesting. You don't need to restructure a 200-person engineering org. You might not have an engineering org at all. Which means you can skip the painful middle step entirely and build the way things are going to work from day one. What does that look like in practice? It means the custom internal tool you've been putting off — the one that's too specific to your business for any vendor to sell you — is now a week-long project, not a six-month one. That integration between your CRM and your inventory system that you've been doing manually with spreadsheets can be automated for the cost of a few dollars a month in AI compute. You can now have a technical strategy without having a technical department. I've been building this way for the past several months, and the speed is surprising — even to me, and I've been writing software for 20 years. Things that I would have estimated at 2-3 months are taking 3-5 days. Not because the problems got simpler, but because the cost of translating a clear problem description into working software has dropped by an order of magnitude. The catch — and there's always a catch — is that you still need someone who knows what to build and why. AI is incredibly fast at writing code. It has no idea whether the code it's writing solves the right problem. The judgment — what to build, how to design it, when it's ready to ship — that's the part that still requires experience. George frames his article around a question every board should be asking: "Which path are we on?" I'd adapt that for anyone running a company that isn't a software company: What's your plan for when the tools you depend on change underneath you — and what are you going to build for yourself that nobody else can sell you? If you're thinking about this and want to compare notes, I'm always up for a conversation!

1

1 Comment

I am super excited this morning to announce Sandgarden's acquisition of Layer! When I first met Jonah Katz and Andrew Hamilton, I fell in love with their enthusiasm and positivity, not to mention what they were building. Two years later, I'm proud to have them join Sandgarden. Thanks to Michael Hirshland for the introduction, and to Andrew and Jonah for believing in what we're building to take a chance. Congrats on getting it over the finish line! Welcome to the team Ayaan G. and Gavyn P.! Finally - check out Doc Holiday! It keeps everyone in the company up-to-date as features ship, starting with release notes and documentation all the way to team-specific Action Plans. Demo video here: https://lnkd.in/grkAuXkh And finally - a shout to Jonathan J. Chou, Nicholle Lamartina, and Tori Stark for the insane help. You never fail to deliver 🥰

108

12 Comments

I thought this was a good and very reasonable talk by Eleanor Millman on how to prioritize projects for a platform engineering team: https://lnkd.in/gaYtUHRH A few insights I took from it: - Unlike product features which can be more directly tied to revenue, platform engineering work is a couple steps removed. But that doesn't mean it's unimportant! Far from it, platform engineering projects can make everyone at the company more efficient, able to produce higher quality software, etc. - While "urgency" is always going to be a factor, you don't want to just be putting out fires, you want be able to prioritize work which is "high impact" ... even if that impact isn't felt immediately. - Good rule of thumb: prioritize "highest impact for lowest effort". - The talk has some ideas on what factors to choose for impact, and how to blend them. For instance "speed of development" is one factor, "cloud cost optimization" might be another. The weighting of different impact factors can change over time, depending on the needs of the business. I would say that a lot of companies don't have much methodology here but I can really see the value in codifying it. You can always change the methodology or the weighting if it's spitting out results that don't pass the smell test or you really feel it is leading the company astray. Clarity can give the org more freedom to put resources behind projects that would otherwise never be taken on. When things are unclear, prioritization still happens implicitly, but the decisions tend to be a lot more random and fear-based and the org is worse off as a result.

14

3 Comments