Richard Knoll

Software agents change coding. AI agents will build the same feature 10 times and select the best version. No limits, no guesswork—just optimized innovation. This is a radical new mindset: iterative, faster than ever, and powered by AI. 🚀 Nikita Shamgunov, CEO of Neon, is focused on the architecture of the future and creating for these next wave of agentic and code-gen applications powered by LLMs. A world where cost, infrastructure, and time barriers are tackled and AI agent development unleashes innovation at scale. The future of software development is parallel, not linear! And we'll need GenAI Friendly serverless database like Neon to make it possible. Watch the full episode to discover: - How serverless Postgres is enabling the next wave of AI-powered apps - Why "breadcrumbs" are crucial for AI agent systems - The architecture behind instant database branching 

1

Mr Clegg is wrong about Open Source — there are very real reasons why it is critical that the source (i.e., data) be provided. Without it, users cannot enjoy the freedoms to use, study, modify, and share the software, and therefore cannot build the next generation of models on it, rather render themselves, their company, and/or country subservient to Meta. They can't even examine the data for issues with security/bias/etc., let alone rectify them. That he confuses copyright for the trademark the Open Source Initiative (OSI) failed to obtain in the 90s suggests he probably shouldn't be commenting on IP issues at all, though I assume they know exactly what they are doing. Folks like Ai2 would like a word about models like #Molmo being "scientific and low-performing" too. And yes, vinyl is great, but I stream everything nowadays. Llama models are indeed very powerful and a great gift to society that counter-balance service-only offerings like OpenAI and Anthropic, but they are clearly not Open Source. Those building on them need to enter the relationship with eyes wide open — if Open Weight model vendors decide to start charging for them, move to service-only, or shift focus then you'll find yourself at a dead-end. "Some critics argue that Meta shouldn’t be using the term “open-source” at all. Current versions of Llama still have restrictions that traditional open-source software doesn’t (including lack of access to datasets). In October, the Open Source Initiative, which coined the term, criticized Meta for “confusing” users and "polluting” the nomenclature, and noted that Google and Microsoft had dropped their use of the term (using the phrase “open weights” instead). Nick Clegg, Meta’s global affairs chief, is blunt in his rebuttal: He says the debate reminds him of “folks who get very agitated about how vinyl is the only true kind of definition of good music.” Only a handful of scientific and low-performing models would fit the definition, he continues: “No one has copyright IP ownership over these two English words.”" #OpenSource #OpenWeight #ArtificialIntelligence #AI #Meta #Llama #EU #EUAIAct

16

5 Comments

1. Gradle Inc. Gradle is like a magic builder for your software projects. Think of building a house - you need to assemble bricks, wood, and other materials. Gradle does this for software. It takes your code, puts all the pieces together, and turns it into a working app or software. It's super fast and can handle big projects with ease, making sure everything is in the right place. 2. Phabricator Phabricator is like an all-in-one project hub for software development. Think of it as a digital workshop where your team can plan, build, and manage projects together. It helps you track tasks, review code, and collaborate in one place. This makes sure everyone stays on the same page, work moves smoothly, and projects are completed efficiently. 3. Gitea Gitea is like a digital storage locker for your code, but with teamwork in mind. Imagine working on a project with friends. You all need to keep track of who did what and make sure you're not messing up each other's work. Gitea helps you do that by letting you save your code, track changes, and collaborate seamlessly. Plus, it’s lightweight and super easy to set up! 4. Huginn Huginn is like your personal internet assistant. Imagine having a smart robot that watches the internet for you and takes action when something happens. Huginn can send you alerts, gather information, and even perform tasks automatically. Whether it’s monitoring news, tracking emails, or managing social media, Huginn makes your online life easier by doing the repetitive work for you. 5. Tabnine TabNine is like an intelligent assistant for coding. Let's say you’re writing a long essay, and someone whispers the next sentence to you. That’s what TabNine does for coding. It predicts what you’re about to type next and offers suggestions, helping you code faster and with fewer mistakes. It’s like having a smart friend who knows exactly what you need. 6. Ansible by Red Hat Ansible is like a remote control for your IT tasks. Imagine having to cook dinner for 100 people in different kitchens. Ansible helps you control all the stoves from one place. It automates tasks like setting up servers, deploying apps, or managing systems, so you don’t have to do it manually. It’s super efficient, reliable, and saves a ton of time. Follow For More 🤖

6

2 Comments

The staging environment trap: Your LLM performs flawlessly during testing. Confidence scores look perfect. Code reviews are green. Then production hits - and you realize confidence scores tell a very different story at scale. If you're leading an engineering team implementing LLMs, you've likely encountered this disconnect between controlled testing and production reality. The challenge isn't just about getting LLMs to work - it's about making them reliably work at scale. Over the next few days, we'll do a technical deep-dive into production-grade LLM validation architectures, specifically examining: 1. Systematic validation beyond basic confidence scores    - Meta-cognitive processing layers    - Recursive self-evaluation patterns    - Real production implementation 2. Multi-layer confidence assessment    - Validation hierarchy design    - Performance impact considerations    - Integration patterns 3. Production-scale implementation    - System architecture requirements    - Performance optimization strategies    - Real-world adaptation patterns This isn't about theory - it's about practical, implemented solutions to real engineering challenges in LLM deployments. Follow us for this technical exploration if you're wrestling with: - Validating LLM outputs at scale - Building reliable confidence assessment systems - Implementing production-grade validation architectures Engineering leaders: What's your biggest challenge in validating LLM outputs in production environments? #AIEngineering #SystemArchitecture #TechnicalLeadership

33

Here are the slides for my GopherCon.EU talk: How to Win Frames and Influence Pointers https://lnkd.in/d7byRQfu #gopherconEU #golang

72

4 Comments

I get asked daily about the "AI bubble" and whether it's real. Are we in an AI bubble? I think to some degree we are in an AI infrastructure bubble, but I also believe It's mini bubble - smaller than many think - and it's growing slower than initially expected. My read is that application and services companies are hitting significant challenges in using the infrastructure they've acquired to build and sell product that customer want. the potential for AI is still very much out there, but addressing specific enterprise-level problems and use cases has been more challenging then initially perceived for many startups. The technology is advancing, but it’s not yet reliable enough for critical business applications. The continued improvements in new versions of GPT is not helping much. Yes, ChatGPT 4o is better, faster, and has awesome voice agent capabilities, but it's mostly a consumer product and doesn't really move the needle for the majority of Enterprise use-cases. These large models are also complicating efforts to build AI systems that excel at specific tasks within enterprises. This kind of progress, while impressive, sometimes seems to diverge from what is needed to boost enterprise adoption. Many investors here in Silicon Valley have mentioned a sense of disillusionment around AI progression. While some of the excitement may be cooling, it’s too early to determine if this will lead to a downturn. If there is a shift, we might see computing, training, and inference costs decrease due to reduced demand. If the hype around AI goes down, more conservative investing could mean less competition for the existing players in the market. How do you see this evolving over the next year or two?

32

16 Comments

I've said before that I think the killer use case for GenAI is not chat or coding but GenVideo, and I'm seeing a shift in the industry as others seem to be making that same realization. I have yet to have a chance to play with Veo 2, but it may very well put OpenAI on a significant defensive.

12

1 Comment

It remains to be seen whether the 'experience' curve commonly observed for mass production (i.e. semi-conductors, electronics,...) would ultimately applied to LLM with subscription as revenue model.

3

Imagine you’re trying to manage a team of 1,000 people, each working on a different task. Now imagine that to give them instructions, you have to call each one individually, deal with line static, and hope they follow your directions exactly. Sounds chaotic, right? That’s what running experiments on a quantum computer used to feel like - until Israeli startup Quantum Machines came along. Quantum computers require precise control. Existing systems rely on analog hardware like mixers and oscillators to communicate with the qubits. But these systems are noisy, need constant tweaking, and can’t handle the scale we need for the quantum computers of tomorrow. Quantum Machines changed the game with their OPX1000. It’s like moving from making individual phone calls to running a group chat with perfect reception. Instead of analog signals, they use Direct Digital Synthesis (DDS), which offers crystal-clear control of qubits, no matter how many. With DDS, everything stays in sync, with almost no noise or signal drift, meaning fewer computation errors and faster results. What does this mean for the future? It means quantum computers can work faster and more accurately. It means scaling quantum systems beyond what we thought was possible. And ultimately, it means quantum computing can move from the lab into solving real-world problems - sooner than we think. And what kind of problems can quantum computing solve? Think of revolutionizing drug discovery by simulating molecular interactions in ways classical computers can’t. Imagine solving some of the toughest climate models, or creating new materials for clean energy,. The promise of quantum computing is nothing short of world-changing. And thanks to Israeli innovators like Quantum Machines, we’re getting closer every day. #QuantumComputing #TechInnovation #FutureTech

12

1 Comment

Update: the epsiode is out: https://lnkd.in/ew89pM4a Tomorrow's podcast episode: How Linear gets so much done, with such a small team. With Sabin Roman, Linear's first engineering manager. 10,000+ companies use Linear, including 66% of Forbes Top 50 AI companies. They are profitable, full-remote and have 25 engineers. Sabin is Linear's first EM - and, as luck would have it, we worked together at Uber for a few years beforehand. We also touch on those parts a bit - comparing what's different at Linear, versus practices at Uber. Search for "The Pragmatic Engineer Podcast" on your favorite player, or subscribe on YouTube to get the episode: https://lnkd.in/eqZqkiqC

141

15 Comments

Stacked Diffs is the best kept secret of Meta and we copied it 💥 Stacking has been months into working at HyperTest. But now this is our default way to build and ship code because it does not block our devs for reviews. Devs love it and also complain why did we not implement this earlier. But why did we need to move to this? The traditional development workflow is this: 1, Finish a feature 2, Submit a PR 3, Await PR approval 4, Merge back into main once approved Once the PR is merged, start work on the next feature. This process takes a long time! The biggest time waster is that PRs sit waiting for review for hours or days. This wait becomes longer if current PR is dependent on a feature in another PR still under review This where stacking comes to rescue Stacking lets devs make many small PRs easily, without waiting for a review. To give a practical example, it’s pretty common that once a dev builds a feature, he wants to build the next feature using the completed one. With stacking, you don’t have to wait for the first feature to be approved and merged; you just keep working! The idea behind stacked diffs is that you can keep working on your main branch, and worry about reviews later. This is why very high performance teams, like the ones Meta, Google, Amazon, Rubrik and others use stacking as the only default way to write and ship code. But stacking is much easier to get started with in smaller teams. The learning curve was negligible but returns are great! We love it. #softwareengineering #stacking #git

16

𝗪𝗵𝘆 𝘄𝗼𝘂𝗹𝗱 𝘄𝗲 𝗹𝗲𝘁 𝗔𝗜 𝗿𝗲𝗽𝗹𝗮𝗰𝗲 𝗽𝗮𝗶𝗿 𝗽𝗿𝗼𝗴𝗿𝗮𝗺𝗺𝗶𝗻𝗴 𝗲𝗻𝘁𝗶𝗿𝗲𝗹𝘆? The idea is floating around: “With tools like GitHub Copilot, why not just skip human pair programming and let AI handle it?” Sure, AI can speed up repetitive tasks, help us get unstuck, and suggest code in an instant. But is that enough to replace pair programming? Not really - and here’s why: • 𝗕𝘂𝗶𝗹𝗱𝗶𝗻𝗴 𝗿𝗲𝗮𝗹 𝘁𝗲𝗮𝗺 𝗱𝘆𝗻𝗮𝗺𝗶𝗰𝘀. Pair programming isn’t just about one person’s productivity - it’s about leveling up the entire team. When two minds collaborate, each coder brings unique insights that grow everyone’s skills, enhancing the team as a whole. AI’s focused on helping the individual; human pairing benefits everyone. • 𝗨𝗻𝗱𝗲𝗿𝘀𝘁𝗮𝗻𝗱𝗶𝗻𝗴 𝗰𝗼𝗻𝘁𝗲𝘅𝘁. AI can be fast and clever, but it doesn’t grasp the big picture - the project’s goals, potential challenges, or the larger context that can impact decisions. A human partner intuitively adjusts to these layers, helping ensure that the code aligns with the long-term vision. • 𝗙𝗼𝘀𝘁𝗲𝗿𝗶𝗻𝗴 𝗼𝘄𝗻𝗲𝗿𝘀𝗵𝗶𝗽 𝗮𝗻𝗱 𝗮𝗰𝗰𝗼𝘂𝗻𝘁𝗮𝗯𝗶𝗹𝗶𝘁𝘆. There’s a big difference between code “suggestions” and true co-ownership. When we pair with another person, we share accountability for the final product. That sense of ownership builds commitment, quality, and trust within a team - something AI can’t replicate. AI is an incredible tool for productivity, but treating it as a “replacement” for pair programming risks missing out on the deeper benefits of human collaboration. What’s your take - could AI ever fully replace pair programming, or are we risking too much by letting go of human-to-human connection? #SoftwareDevelopment #Collaboration #PairProgramming #Copilot #AI

8

AI summaries of customer feedback are boring and useless The problem with AI summaries is… that they summarize. Averaging everything you’re summarizing into a boring grey goo For example, feeding Spotify app store reviews to GPT gives you something like: Spotify app reviews reveal a mix of praise and criticism. Positive Feedback: • Music Discovery: Users appreciate Spotify's music discovery features and recommendation algorithms • User Interface: The interface is praised for its ease of use, making it simple to find songs and create playlists Negative Feedback: • Ads and Free Tier Limitations: Users on the free tier report frequent ads which can be frustrating  • Subscription Issues: Some users experience difficulties managing subscriptions • Content Recommendations: Some users feel the algorithm suggests unwanted genres This is boring/ useless… The purpose of a summary is to hit highlights, not to homogenize everything. A better version might look like: 𝐔𝐬𝐞𝐫𝐬 𝐚𝐫𝐞 𝐡𝐚𝐯𝐢𝐧𝐠 𝐢𝐬𝐬𝐮𝐞𝐬 𝐰𝐢𝐭𝐡 𝐬𝐡𝐮𝐟𝐟𝐥𝐞 𝐦𝐨𝐝𝐞 𝐚𝐧𝐝 𝐬𝐨𝐧𝐠 𝐪𝐮𝐞𝐮𝐞 (𝐫𝐞𝐩𝐨𝐫𝐭𝐞𝐝 𝐛𝐲 𝟒𝟎𝟎 𝐮𝐬𝐞𝐫𝐬 𝟒% 𝐨𝐟 𝐭𝐨𝐭𝐚𝐥) "𝑟𝑒𝑐𝑒𝑛𝑡𝑙𝑦 𝑖𝑛 𝑠ℎ𝑢𝑓𝑓𝑙𝑒 𝑚𝑦 𝑠𝑜𝑛𝑔𝑠 𝑤𝑖𝑙𝑙 𝑓𝑟𝑒𝑒𝑧𝑒 𝑎𝑓𝑡𝑒𝑟 𝑎 𝑓𝑒𝑤 𝑠𝑘𝑖𝑝𝑠"  "𝐼 𝑚𝑎𝑘𝑒 𝑝𝑙𝑎𝑦𝑙𝑖𝑠𝑡𝑠, ℎ𝑖𝑡 𝑡ℎ𝑒 𝑝𝑙𝑎𝑦 𝑏𝑢𝑡𝑡𝑜𝑛, 𝑎𝑛𝑑 𝑡ℎ𝑒 𝑠𝑎𝑚𝑒 4 𝑠𝑜𝑛𝑔𝑠 𝑝𝑙𝑎𝑦 𝑜𝑣𝑒𝑟 𝑎𝑛𝑑 𝑜𝑣𝑒𝑟 𝑖𝑓 𝐼 𝑑𝑜𝑛’𝑡 𝑟𝑒𝑠ℎ𝑢𝑓𝑓𝑙𝑒 𝑚𝑦𝑠𝑒𝑙𝑓" "𝑇ℎ𝑒 𝑞𝑢𝑒𝑢𝑒 𝑖𝑠 𝑏𝑟𝑜𝑘𝑒𝑛 𝑓𝑜𝑟 𝑚𝑒. 𝐴𝑑𝑑𝑒𝑑 𝑠𝑜𝑛𝑔𝑠 𝑎𝑟𝑒 𝑟𝑒𝑚𝑜𝑣𝑒𝑑" 𝐈𝐬𝐬𝐮𝐞𝐬 𝐰𝐢𝐭𝐡 𝐜𝐚𝐫 𝐦𝐨𝐝𝐞 𝐦𝐚𝐲 𝐛𝐞 𝐜𝐚𝐮𝐬𝐢𝐧𝐠 𝐬𝐚𝐟𝐞𝐭𝐲 𝐢𝐬𝐬𝐮𝐞𝐬 (𝐫𝐞𝐩𝐨𝐫𝐭𝐞𝐝 𝐛𝐲 𝟖𝟎 𝐮𝐬𝐞𝐫𝐬 𝟎.𝟖% 𝐨𝐟 𝐭𝐨𝐭𝐚𝐥) “𝑁𝑜𝑤 𝑒𝑣𝑒𝑟𝑦 𝑡𝑖𝑚𝑒 𝐼 𝑛𝑒𝑒𝑑 𝑡𝑜 𝑓𝑖𝑥 𝑚𝑦 𝑒𝑞𝑢𝑎𝑙𝑖𝑧𝑒𝑟 𝑐𝑢𝑧 𝐼’𝑚 𝑜𝑛 𝑡ℎ𝑒 ℎ𝑖𝑔ℎ𝑤𝑎𝑦 𝑎𝑛𝑑 𝑤𝑎𝑛𝑛𝑎 𝑡𝑢𝑟𝑛 𝑢𝑝 𝑡ℎ𝑒 𝑣𝑜𝑙𝑢𝑚𝑒 𝐼 ℎ𝑎𝑣𝑒 𝑡𝑜 𝑑𝑖𝑠𝑎𝑏𝑙𝑒 𝑐𝑎𝑟 𝑚𝑜𝑑𝑒 𝑓𝑖𝑟𝑠𝑡? 𝐻𝑜𝑤’𝑠 𝑡ℎ𝑎𝑡 𝑠𝑎𝑓𝑒𝑟" “𝐼 𝑤𝑖𝑠ℎ 𝑡ℎ𝑒 𝑐𝑜𝑛𝑡𝑟𝑜𝑙𝑠 𝑤𝑜𝑢𝑙𝑑 𝑠𝑡𝑎𝑦 𝑖𝑛 𝑡ℎ𝑒 𝑠𝑎𝑚𝑒 𝑜𝑟𝑖𝑒𝑛𝑡𝑎𝑡𝑖𝑜𝑛 𝑏𝑒𝑡𝑤𝑒𝑒𝑛 𝑟𝑒𝑔𝑢𝑙𝑎𝑟 𝑎𝑛𝑑 𝑐𝑎𝑟 𝑚𝑜𝑑𝑒” 𝐓𝐡𝐞𝐫𝐞 𝐢𝐬 𝐚 𝐬𝐩𝐢𝐤𝐞 𝐢𝐧 𝐢𝐬𝐬𝐮𝐞𝐬 𝐰𝐢𝐭𝐡 𝐨𝐟𝐟𝐥𝐢𝐧𝐞 𝐦𝐨𝐝𝐞 (𝟒𝟎𝟎 𝐮𝐬𝐞𝐫𝐬, 𝟒% 𝐨𝐟 𝐭𝐨𝐭𝐚𝐥 𝐮𝐩 𝟏𝟎𝐱 𝐌𝐨𝐌) "𝐷𝑜𝑤𝑛𝑙𝑜𝑎𝑑𝑒𝑑 𝑠𝑜𝑛𝑔𝑠 𝑎𝑟𝑒 𝑢𝑛𝑎𝑣𝑎𝑖𝑙𝑎𝑏𝑙𝑒 𝑤ℎ𝑒𝑛 𝑜𝑓𝑓𝑙𝑖𝑛𝑒, 𝑒𝑣𝑒𝑛 𝑡ℎ𝑜𝑢𝑔ℎ 𝑡ℎ𝑒𝑦 𝑠ℎ𝑜𝑤 𝑎𝑠 𝑑𝑜𝑤𝑛𝑙𝑜𝑎𝑑𝑒𝑑."  "𝐴𝑝𝑝 𝑘𝑒𝑒𝑝𝑠 𝑑𝑒𝑓𝑎𝑢𝑙𝑡𝑖𝑛𝑔 𝑡𝑜 𝑜𝑓𝑓𝑙𝑖𝑛𝑒 𝑚𝑜𝑑𝑒 𝑒𝑣𝑒𝑛 𝑤ℎ𝑒𝑛 𝐼 ℎ𝑎𝑣𝑒 𝑖𝑛𝑡𝑒𝑟𝑛𝑒𝑡 𝑐𝑜𝑛𝑛𝑒𝑐𝑡𝑖𝑜𝑛." This is a sampling, not a summary of the most interesting specific comments made by users, including specific facts and statistics. Moreover it is actionable and gives a starting point as to where we might dig further. This is not something AI is good at out of the box. It requires the ability to quantify what is being reported by users, decide on the most important points and present an output that is suitable to the use case Others see the same?

34

3 Comments

Ollama Workbench just got a MAJOR upgrade! Here's what's new: 🔨 Build: Our autonomous multi-agent software development system will blow your mind. Imagine building a whole project just by describing it! Build leverages specialized AI agents to handle everything from planning and coding to testing and documentation, completely revolutionizing your software development process. ⛓️ Nodes: Our new visual workflow builder gives you the power to create custom AI processing pipelines. Chain together multiple models, inputs, and outputs with ease, and fine-tune each node's settings for complete control. ☁️ External Providers: This is HUGE! Now you can seamlessly integrate OpenAI, Groq, and various search providers directly into Ollama Workbench. Can't run Ollama locally? No problem! Leverage the speed of Groq, the quality of OpenAI, and the comprehensive search capabilities of Google, Bing, SerpApi, and more. Real-World Problem Solvers: - Scenario 1: Need to research a complex topic? The "Research" feature now uses multiple AI agents, each with its own search strategy and library, to generate a detailed report with references, saving you hours of manual work. - Scenario 2: Struggling with brainstorming new ideas? Brainstorm mode lets you collaborate with a team of AI agents, each with unique perspectives and memories of past sessions. The result? More creative, out-of-the-box thinking. - Scenario 3: Want to build a simple web app without writing any code? Use the "Build" feature to describe your app (e.g., "a Streamlit app that shows current weather data"), and let the AI agents do the heavy lifting! Ollama Workbench keeps getting better. Try it out today and see how it can empower your AI journey! You realize it's free, right? So, what's stopping you? https://lnkd.in/eNft2Pje #Ollama #LLM #AI #Workflow #Automation #SoftwareDevelopment #Productivity #OpenAI #Groq #Search #Research #Brainstorm #Innovation

2

2 Comments

OpenAI just announced their o3 model (they skipped the o2 name). We'll see it next year. Software development will change significantly in the upcoming year. It already has, but that's just the beginning. And not only software development... https://lnkd.in/ehXeJtss

47

AMAZON??? This is again why most CEO are unfit to hold the job? Investors hired you to bet on a startup that is not successful because you idiots don’t understand that acquiring startup that is not able to sell the product is called gambling . And by the way what GenAi can do for you. We can’t sell GenerativeAI . We have to find the idea . The market then may be there is room or case to enhance our revenues. You don’t bet if you don’t see the money then hire who can see or resign, why are you getting this huge compensation to run the company 😁 https://lnkd.in/gf3venXi

2

5 Comments

OpenAI recently acquired Rockset, joining a lineage with roots in Google and Facebook. Rockset originated from an open-source Facebook project called RocksDB, which itself was based on Google's open-source LevelDB. So, what's the acquisition about? Products like ChatGPT can be made even more relevant for your use case by giving them real-time access to business data. And that's where a database like Rockset comes into picture. Rockset allows real-time data ingestion from a variety of sources at play in today's businesses, and provides fast search & analytics on top, for LLM-powered products like ChatGPT. **** Book a 1:1 with me https://lnkd.in/gdQ7mY5f #openai #llm #rag #rockset #rocksdb #leveldb #google #facebook #ai

5

1 Comment