About
I've been writing code since I was a kid, studied bioinformatics when applying computer…
Articles by Dror
Activity
21K followers
-
Dror Berman reposted thisDror Berman reposted thisI spent decades building companies where the constraint was always hiring enough engineers fast enough. That constraint is gone. Dror Berman and I wrote this because we think most people are underestimating what is happening. Not the technology. What it makes possible. https://lnkd.in/exyGcTtP
-
Dror Berman reposted thisDror Berman and I wrote about what changes when agents become the primary consumers of technology. This piece from Bauplan takes that idea further: when agents are the main consumers of software, the advantage of a flashy interface disappears. Codebases become disposable. What really matters is data and infrastructure.Dror Berman reposted thisDror Berman and Eric Schmidt recently wrote about what they call the agent economy. Their argument: when agents are both the builders and consumers of technology, the constraints that shaped everything we built for humans fall away. We read it and asked ourselves: what does this mean for infrastructure? Platforms built their moats on human interfaces. Dashboards, consoles, configuration wizards. The markup on compute was justified because a human needed the platform to do their job. Agents don't need any of that. They need APIs, branching, transactional guarantees, and pricing that doesn't collapse at agent-scale volume. We wrote a response exploring what happens to the software stack when the interface layer dissolves, where value accrues next, and what infrastructure must look like to serve agents as first-class operators.
-
Dror Berman shared thisFor as long as technology has existed, it has been built by humans, for humans. We still measure whether a cancer drug is working by having a radiologist measure tumor diameter, a single number, because that is what a human can interpret. That era is ending. Eric Schmidt and I were struck by how fast the ground is shifting, and writing this together sharpened our thinking on where it leads. The constraint is no longer what we can build. It is what we can imagine. And for the first time, we are building things that surpass our ability to fully understand them.
-
Dror Berman shared thisWe live in crazy times. Startups with 10 people are shipping what used to take hundreds. Most companies respond in one of two ways: adding AI to existing workflows, or making incremental org changes rather than thinking from first principles about how the company should be organized. Both miss the point. The right starting point is a simple question. For every role: what decisions does this person make? If the answer is mostly none, that role needs to be redesigned around judgment, direction, and ownership. Most companies aren't doing this. So instead of people owning more, they spend their time reviewing agent output. That's not directing. That's babysitting. We're moving from companies built around headcount to companies built around decision-makers. Fewer people, each one generating dramatically more value. That's not a cost-cutting story. That's a leverage story.
-
Dror Berman shared thisThe product manager role is being redefined. Over the past several months, I've been talking to product leaders at companies ranging from early-stage startups to LinkedIn and Rippling. I wanted to understand: what's actually changing? AI now generates specs, PRDs, research summaries, personas, UX flows, and code scaffolds. What used to take days takes minutes. But here's what most people are missing: the value of a PM is no longer in producing those artifacts. PMs are now expected to prototype and build themselves. This is the new "PMs must know SQL" moment that changed the role a decade ago. My friend Effi Fuks Leichtag, CPO at Next Insurance, said it best: "For PMs, it's heaven. Your job is now to manifest something profound, so people can say: heck, this is great, let's go for it." The shift is clear: the winners will be those who move their value from producing documentation to producing insight, judgment, and direction. Everything else is being automated.
-
Dror Berman shared thisWhen someone tells you they want to destroy your country, believe them. For 47 years, the Iranian regime has chanted "Death to America" and "Death to Israel." Not as a metaphor. Not as political theater. As a mission. They built a nuclear program around it. They amassed thousands of ballistic missiles. In January they tested an ICBM with a 10,000 kilometer range. That puts Washington, D.C. within reach. We should've learned this lesson already. Before September 11th, Al Qaeda told us exactly what they planned to do. Before October 7th, Hamas made no secret of what they wanted. In both cases, we didn't act in time. The cost of that inaction was measured in thousands of lives. The U.S. and Israel are right to take action. This is not about politics. It's about taking a regime at its word. A regime that authorized the development of nuclear warheads. That armed for decades thousands of terrorists in the Middle East. That has spent nearly five decades exporting terror while strangling its own people. But what really excites me is what comes after. The Iranian people are not the Iranian regime. 70 to 80 percent of Iranians say they would not vote for the Islamic Republic. The protests that erupted in December are the largest since the 1979 revolution itself. These are not fringe movements. This is a population of 90 million, young and highly educated, that has been demanding a secular democracy for years. And when that happens, the economic transformation will be staggering. I see it every day in Silicon Valley. Some of the most extraordinary founders, engineers, and business leaders in the valley are Iranian. Dara Khosrowshahi helped build Uber into a $150 billion company. Ali Ghodsi built Databricks into one of the most valuable private companies in the world. Pierre Omidyar created eBay. Iranian Americans hold more advanced degrees per capita than any other ethnic group in the United States. That is what happens when a fraction of Iran's talent gets a chance to build freely. Now imagine what happens when 90 million people get that same chance. A young, educated population sitting on massive natural resources, at the crossroads of Asia, Europe, and the Middle East. The only thing standing in the way has been the regime itself. This is a country with 5,000 years of civilization. A country that produced some of the greatest mathematicians, poets, and scientists in human history. A country whose people have been oppressed, sanctioned, and isolated by a government that hijacked their future 47 years ago. The regime told us who they were. We should believe them. And the Iranian people have been telling us who they are too. In the streets, in exile, in the companies they have built around the world. It is time to listen.
-
Dror Berman shared thisSome founders you'd back even if they pitched a lemonade stand. Deyan Vitanov is one of those people. He's one of the sharpest founders I've worked with. He has a clear vision of where the world is going, but stays genuinely open to feedback and new information. That combination is rare. Today Deyan and Petar launched Jest out of stealth. We led their $7M seed round. Here's what got me excited: AI is making it dramatically easier to build new products. Games included. When building gets infinitely easy, the bottleneck shifts to distribution. And right now, distribution is broken. Apple and Google built walled gardens that charge developers 30% for the privilege of being found. For indie game studios, that's often the difference between surviving and shutting down. Jest is attacking this head on. A marketplace for messaging games that lets people play instantly inside their texts, no download required. Developers keep 90% of revenue. And with the Jest Games Fund, they're deploying up to $1M per title to help studios build the first generation of messaging native games. The next wave of apps won't live in the app store. They'll live where people already are. In conversation. Proud to back Deyan and Petar.
-
Dror Berman shared thisMost companies are adding AI to workflows that shouldn't exist anymore. Capital, time, and expertise used to constrain what you could build. All three have loosened at once. That changes everything. How you start a company. How you run it. How you decide what to build and how you take it to market. Most of the companies I work with that aren't AI-native are grappling with this. They start from what they have and try to introduce tools and efficiency. That's the wrong instinct. The right move is to rethink each workflow and function from scratch. Not "how do we make this faster?" but "if we were starting today, how would we build this?" I've spent the last six months talking to product leaders and founders about how they're reorganizing. The pattern is clear: the companies moving fastest didn't start by adding AI. They started by questioning every assumption about how work gets done in light of what AI unlocks. Over the next few weeks, I'll share what I'm seeing from working closely with this next generation of AI-native companies, and from conversations with leaders trying to make this shift. The ground is moving, fast.
-
Dror Berman shared thisMost AI labs are trying to make the same architecture faster. Stefano and Volodymyr bet on replacing it entirely. When we backed Inception at Innovation Endeavors, the prevailing view was that autoregressive was the only path for language models. Stefano saw something different: diffusion, the same approach behind image generation, could unlock a fundamentally faster way to reason. Mercury 2 is the proof. 1,000 tokens per second. 5x faster than leading reasoning models. Not through optimization tricks, but through a different architecture altogether. This is the kind of bet we love. Congrats Stefano, Volo and the entire Inception team!Dror Berman shared thisOur team has spent years working on a bet: that diffusion could unlock fundamentally faster language generation than autoregressive models allow. Today we're launching Mercury 2, the fastest reasoning LLM and first reasoning dLLM, proving that bet was worth making. Mercury 2 delivers 5x faster performance than leading speed-optimized LLMs, achieving >1,000 tokens per second throughput. That speed unlocks what’s been missing in production reasoning: Multi-step agents without delays. Interactive search and voice assistants with tight latency budgets. Long-form vibe coding with instant results. The key is diffusion-based generation. Instead of committing one token at a time like a typewriter, Mercury 2 generates like an editor: parallel, iterative refinement. It can catch its own errors during generation, maintain better control, and do it all with dramatically lower inference cost. Huge thanks to our whole team at Inception, who turned years of research into reality, and to our partners who believed in this vision when it was just equations on a whiteboard
-
Dror Berman liked thisDror Berman liked thisJust wrapped up an incredible whirlwind trip in Singapore with our AlphaSense team, and I’m still processing how energizing the experience was. From back-to-back 1:1s and team meetings to “power breakfasts”, office Gong christenings, unforgettable AMAs and dinners, every moment reinforced just how special this group is. The Singapore revenue organization is made up of some of the most confident, sophisticated, and driven people I’ve had the privilege to spend time with. What stood out most wasn’t just the extraordinary talent, it was the collaboration, the ambition, and the pride everyone brings to building something meaningful together. I’m deeply grateful for the warm hospitality and the effort that went into making the visit so impactful. Truly blown away. If this week was any indication, this is only the beginning. The foundation here is exceptional, and I have no doubt we’re going to continue to see explosive growth in APAC over the next few years. Huge thank you to the entire Singapore team (too many to mention!), already looking forward to the next visit.
-
Dror Berman liked thisDror Berman liked thisAt Skylo, we believe that ‘one bar’ should be a thing of the past. Tami Erwin, Justin Springham and I discussed how our relentless focus on enhancing customer experience is shaping Skylo’s network and service roadmap. Elizabeth Starr Miller Coyne eloquently sums up our perspective at MWC: The future of satellite isn’t about putting more hardware in orbit. It’s about making space connectivity invisible, standardized and seamlessly integrated into the networks people already use every day. Read our conversation in Fierce Network here: https://lnkd.in/gAiWNNuu #MWC26 #StandardizedSky
-
Dror Berman reacted on thisDror Berman reacted on this"When the end consumer of technology is no longer a human, the output no longer needs to be constrained by what a person can sense, comprehend, or attend to" I keep coming back to this article co-authored by one of my former investors, Dror Berman (link in comments). It's provocative but incredibly thoughtful and I think it outlines the significance of the moment in front of us. During a catalyst phase (think the advent of smartphones) so much of the world changes at once as resources that were scarce become commoditized. When the world looks different, there's an opportunity to build for the new ecosystem. In regulated environments like healthcare, it will be interesting to see how quickly the FDA embraces (or doesn't) the direction we're inevitably headed.
-
Dror Berman liked thisDror Berman liked thisAugment Code just embarrassed every AI coding tool using the exact same AI. SWE-Bench Pro results, same Claude Opus 4.5: Augment: 51.80% Cursor: 50.21% Claude Code: 49.75% Same model. Different outcome. The secret: Context Engine. While other tools keyword-search your codebase like it's 2015, Augmentcode builds a semantic map of 400,000+ files understanding why code exists, not just where. Real bug that broke competitors: Deep BCrypt issue in Ansible. Others found the surface API and stopped. Augmentcode traced relationships 6 layers down and fixed it. The flex: They open-sourced it as an MCP server. Plug it into Cursor, Claude, whatever 70% performance boost. Spotify, MongoDB, Webflow already use it. link: augmentcode.com
-
Dror Berman liked thisDror Berman liked thisMost CTOs are burning weeks testing AI tools that don't improve their engineering org. I've evaluated hundreds. Only 5 made it into my actual weekly workflow. Here's what they do: → Code review that sees your full architecture not just the diff → Orchestrates multiple AI agents across cloud environments from a single editor → Reasons across your entire codebase, not just the file you're editing → Compresses hours of reading documentation into minutes → Turns Slack conversations into tracked tasks instantly without leaving the thread Together they help me: → increase engineering output without increasing headcount → keep tasks from falling through the cracks during fast-moving sprints → reduce low-leverage engineering time → maintain velocity without sacrificing code standards I put together a short breakdown of exactly which tools, how I use each one, and the failure modes to avoid. Comment "Mesa" and I'll send it over.
-
Dror Berman liked thisDror Berman liked thisNice to see LightTable mentioned in this Andreessen Horowitz piece. AEC has been overlooked by the broader tech world for a long time, despite how much friction exists in the process. Feels like that's starting to change. https://lnkd.in/ezydmfj8Every Building You've Ever Been In Was Designed By Software Built in 1997Every Building You've Ever Been In Was Designed By Software Built in 1997
-
Dror Berman liked thisDror Berman liked thisToday, we're releasing PoliTax Split — a document splitting benchmark we built by hand-labeling 74 tax filings publicly released by the last four White House administrations. Document splitting is a deceptively hard task, where even the best models score 64.1% F1 and 50% recall. For many of our customers in finance, real estate or healthcare, document splitting is often the first step of their pipelines. When ingesting 200+ page mortgage or insurance packages, splitting is often the first step to make the context more manageable for downstream processing. Document splitting is a hard task, requiring classification + localization, done in bulk, across hundreds of pages. We've found that even the best frontier LLMs struggle with it. Raw models have decent precision (they rarely split incorrectly), but struggle with recall (they miss large numbers of real boundaries). The best raw model (Gemini 3.1 Pro) scored 64.1% F1 and 50% recall (meaning half of all real boundaries were missed). We built multiple internal benchmarks to pinpoint these failure modes and drive the development of Extend's splitter solution to close the gap where frontier models still struggle. When running the same models through Extend's splitting harness, recall jumped 17–44 points for every single model we tested. - Opus 4.6: 34.8% → 73.8% - GPT-5.4: 31.7% → 67.6% - Gemini 3.1 Pro: 50.3% → 71.4% Our harness works by controlling what the model sees and how it reasons: 1. We parse the document with layout and OCR models 2. Manage context across smaller chunks so boundary information isn't lost 3. Reconcile split decisions across the full document in multiple passes The LLM can focus purely on reasoning through the context and instructions, and not on fighting the PDF. Extend's splitter implementation not only improves performance, but also splits incredibly fast. As an example, a customer recently sent a 1,044-page loan document through our splitter. It found 315 sub-documents in 4 minutes and 43 seconds(0.27 seconds per page). We're publishing our full learnings, covering how models perform using their native pdf input, page images, and operating over parsed markdown pages. Link to the full benchmark below.
-
Dror Berman liked thisDror Berman liked thisI now wake up early to spend time with Anthropic Claude. I'm multiplying myself, and building agents that work for me 24 hours, analyzing things for me, preparing me for my day, and more. I have a huge unfair advantage when I compare myself now to myself a year ago. When I'm at home I can't stop. I develop games for my kids that teach them Math, and Hebrew, and enable them to think about investing, donation and giving back to community. For the life of me, I don't know how can anyone spend time on TikTok when they can build code instead. What a great time to be alive. No limits. Execution is meaningless. Creativity, hustle, data, distribution ....
-
Dror Berman reacted on thisMind-blowing!
Experience & Education
-
Innovation Endeavors
View Dror’s full experience
See their title, tenure and more.
or
Volunteer Experience
Languages
-
English
Native or bilingual proficiency
-
Hebrew
Native or bilingual proficiency
Other similar profiles
Explore more posts
Explore top content on LinkedIn
Find curated posts and insights for relevant topics all in one place.
View top content