Daniel Waisberg

CedarDB is incredibly cool and more people should know about it. They’re a team of PhDs in Munich building a new relational database, on top of almost 10 years of academic research, that crushes existing benchmarks and maybe (finally?) gets us to the HTAP grail. The core idea is that existing RDBMSes like MySQL and Postgres were built more than 30 years ago, on assumptions about hardware constraints that are just not true anymore. These ecosystems have evolved admirably but ultimately…it’s a database. It’s built not to change very much. Here are a few of the ways that CedarDB is rethinking every element of the database: 1) A better query optimizer In the last 30 years we’ve made a lot of progress on how to optimize SQL queries, to the point where an optimized query can easily outperform a not-so-optimized query by a ton. CedarDB did a few things on this front, implementing the unnesting algorithm developed by Thomas Neumann (one of the leaders of the Umbra research project CedarDB came from), developing a novel approach to join ordering using adaptive optimization that can handle 5K+ relations, and some more stuff. 2) What if your database was actually a compiler? CedarDB doesn’t interpret queries, it instead generates code. For every SQL query that a user writes, CedarDB processes, optimizes it, and generates machine code that the CPU can directly execute. This has been a holy grail for a while, and they implemented it via a custom low-level language that is cheap to convert into machine code. 3) Taking advantage of all cores / Ahmdal’s law Distributing fairly between all available cores is notoriously difficult. Their clever approach to this problem is called morsel-driven parallelism. CedarDB breaks down queries into segments: pipelines of self-contained operations. Then, data is divided into “morsels” per segment – small input data chunks containing roughly ~100K tuples each. 4) Rethinking the buffer manager The idea of the revamped buffer manager in CedarDB is that you can (and should) expect variance not just in data access patterns, but in storage speed and location, page sizes and data organization, and memory hierarchy. CedarDB’s buffer manager is designed from the ground up to work in a heavily multi-threaded environment. It decentralizes buffer management with Pointer Swizzling: Each pointer (memory address) knows whether its data is in memory or on disk. 5) Building a database for change CedarDB’s storage class system employs pluggable interfaces where adding new storage types doesn’t require rewriting other components. E.g. if CXL becomes the go-to storage interface at some point in the future, you don’t need to write another whole component, you just need another endpoint for the buffer manager. Anyway these are just a few of the ideas they’re bringing to the table. Maybe it’s because they’re in Germany, maybe it’s because they’re just really humble, but more people should know about this team!!

123

4 Kommentare

What are the Key Challenges In Agentic AI? #AgenticAI isn’t risky. But autopilot thinking is. The Problems Are Real. But So Are the Fixes. Agentic AI isn’t a threat to manage. It’s a capability to mature. And maturity comes from alignment - not acceleration. So before you ask, “How do we make AI faster?” Ask, “How do we make it wiser?” Here’s the part most people miss: These aren’t warnings. They’re invitations to build better systems, teams, and habits. Agentic AI isn’t dangerous because it acts alone. It’s dangerous when we stop paying attention. Here’s how to stay in charge - without slowing innovation down: #EthicalOversight → Make ethics visible. Don’t treat bias or privacy like fine print. Track them like you track KPIs — in dashboards, not in disclaimers. #TaskValidation → Add checkpoints. If AI can explain what it did and why, you can trust its process — not just its outcome. #ToolReliability → Plan for failure. Every automation chain has a weak link. Expect things to break — and design graceful recovery paths. #HallucinationControl → Verify before you amplify. AI can sound confident while being wrong. Require evidence. Build “show your source” habits. #ContextRetention → Give it memory with meaning. AI remembers data, not purpose. Help it keep the why, not just the what. #OutcomeDrift → Define success like a human, not a spreadsheet. Don’t just ask, “Did it complete the task?” Ask, “Did it make progress that matters?” #HumanAICollaboration → Build shared language. The smartest AI still needs translation. Train humans to ask better — and AI to explain clearer. #ResourceOptimization → Automate with discipline. An ungoverned agent burns time, compute, and cash. Orchestrate workflows. Don’t just unleash them. #RealTimeReasoning → Teach AI to adapt, not just react. In dynamic environments, rigidity breaks. Combine AI logic with live context feeds. #SafetyAlignment → Keep values on a feedback loop. Ethics aren’t one-and-done. Build systems that flag when behavior drifts from intent.

4

CLAUDE AI: RE-VERSIONING TIP - Claude AI fails to update the filename of an artifact after re-versioning. [TLDR] Here is the remedy... REMEDY: Prompt Claude: "Before regenerating the code, first establish the updated filename of the artifact. Each updated version number must be contiguous." Place that warning towards the beginning of your prompt, NOT at the end. Claude must first establish the correct filename BEFORE it executes the task. OBSERVED ISSUE: Claude fails to update the version number in the filename of the downloadable artifact. It changes the version number INSIDE the artifact, but does not create a new version number in the filename itself (e.g. "v1._x_x"). If you don't correct the issue, your versions won't be contiguous in Claude's memory. ROOT CAUSE: Claude has a blindspot: it can't verify the filename output, and it can't modify it correctly UNLESS... it regenerates the entire artifact. COROLLARY ISSUES: Claude wastes your tokens with unnecessary regenerations, and you get massive headaches trying to get the machine to create the correct version number. WHAT THIS PROMPT WILL SOLVE: - No more headaches. - Clean, continuous versioning of the filenames in Claude's memory - Preventing a huge waste of tokens due to useless artifact regeneration - Helping #Anthropic preserve their computing resources #claudeai #reversioning #aibestprompt #filenamingconventions #edenads #generativeengineoptimization

1

YAPASL - Yet Another Post About Semantic Layers... sorry (not sorry). I’ve been doing the BI thing for a long time, and this post from Serge Gershkovich at our partner SqlDBM and our own Elena Goryainova really hit on something I’ve seen throughout my career: https://lnkd.in/gzMQDjt2 One of the hardest parts of BI has never been the dashboards or the SQL — it’s getting everyone aligned on what the data means. You can have great pipelines and beautifully modeled tables, but if every team defines metrics a little differently, you end up with multiple “truths.” And once trust in the numbers erodes, everything downstream gets harder. And now, with AI in the mix, that challenge is amplified. AI systems don’t magically understand your business logic. They don’t know your definitions, your joins, your rules of thumb. So if BI tools interpret data one way and your AI assistant interprets it another, users understandably ask, “Which answer do I trust?” This is why semantic models are becoming such a critical foundation. They let you define meaning once — and share it across both BI and AI. No more dueling definitions, no more mismatched assumptions. Snowflake’s Semantic Views lean into exactly this idea: a single semantic layer that works across SQL, dashboards, and conversational analytics so everything speaks the same language. If you’ve run into the “three versions of the truth” problem (and who hasn’t?), the blog is worth a read.

69

3 Kommentare

Amazon is suing Perplexity over its browser's automated shopping feature This was bound to happen We're watching the early days of agentic commerce unfold—AI that doesn't just recommend products but actively shops, compares, and buys on your behalf. Every major platform shift faces this moment. Travel booking sites disrupted airlines. Price comparison tools challenged retailers. Aggregators forced banks to open their APIs. The pattern is consistent: incumbents resist, users demand convenience, and eventually the market adapts. My take: Users will win this one too. Amazon will adapt, not because they want to, but because they'll have to. The future isn't about controlling the transaction, it's about winning the customer through better prices, service, and value.

4

1 Kommentar

smart bit from Jason Cincotta's post: https://lnkd.in/edk9AKmT "In the AI world, guests don't see 10 blue links. They see a single conversational answer. LLMs don't rank pages. They synthesize from sources they can parse reliably."

2

5 Kommentare

The future of search isn’t search. It’s answers, summaries, and suggestions from LLMs. I gave a talk last week on how we’re scaling Fonzi’s search visibility with Generative Engine Optimization (GEO). We covered: ⚡ What GEO really is ⚡ How to rank in LLMs ⚡ How we’re building a scalable content system at Fonzi ⚡ Tactical examples you can use If you’re thinking about content in a world where ChatGPT, Perplexity, and Google SGE dominate results, give it a watch. See the first comment for the link. And join the conversation! What are you working on for #GEO?

22

2 Kommentare

I’ve noticed that ChatGPT has a new feature: when you ask it for factual information, it first searches the web . If it doesn’t find what you’re looking for, it then asks whether it should look in more specific sources, like books, interviews, or YouTube videos ( other tiers of the web in its hierarchy ). If that still yields nothing, it offers to expand the search further into forums, academic journals, or other rarer sources etc While this may seem interactive, it actually feels quite unnatural by modern search standards. In today’s systems, it’s far more common to have multi-tier retrieval architectures, for example, a very fast layer covering the top 100 million documents, another slower but broader layer searching several billion, and specialized retrievers for media types like videos or other „documents“ . These systems automatically and intelligently expand the scope of the search when needed, using learned optimization logic. By contrast, ChatGPT’s approach, asking the user each time whether it should “search in forums” or “look in rare sources”, feels so outdated in 2025. A well-designed retrieval system should handle that expansion automatically, without interrupting the user flow.

13

3 Kommentare

We added an llms.txt file to TripWays.com, and an interesting signal showed up in search behavior. On December 24, we deployed llms.txt to clarify how AI systems and crawlers should interpret our inventory, structured data, and booking pages across a large marketplace. Within 48 hours, Google Search Console showed a noticeable change, not in impressions or rankings, but in how results were rendered and clicked. Clicks increased while impressions stayed flat, driven primarily by a lift in: • Product snippets • Review snippets What stood out most was where the lift came from. There was no single “winning page” or breakout query. Instead, many pages picked up small gains (+2, +3 clicks), which compounded across a large inventory. That pattern usually points to a site-wide interpretation shift rather than a ranking spike. Our working hypothesis is that clearer machine-readable signals helped Google trust and render our structured data more consistently across the catalog, especially important for large, dynamic marketplaces. Early signal, not a conclusion, but a useful reminder that as search and discovery become more AI-mediated, clarity for machines increasingly translates into clarity for users. We’ll continue to monitor over the coming weeks and share our learnings as they evolve.

9

We didn’t just add AI. We added it where it actually belongs. moara.io just released three Claude-powered integrations built from weeks of listening to real researchers and refining what truly improves workflow. Agentic search. Custom enrichment fields. Library wide chat. This is what a modern reference manager should look like. 🚀

2

We launched 𝗣𝗗𝗔𝗥𝗣 two days ago. The inbound hasn't stopped since. Protocols, institutional desks, integration partners, builders - the volume of genuine interest told me more than any metric could. Now let me tell you what it actually took to build it. When Timm Reinsdorf, Nadine Wilke and I decided to move Particula's risk ratings onchain, the product vision was clear. The engineering reality was not. 𝗥𝗶𝘀𝗸 𝗱𝗮𝘁𝗮 𝗶𝘀 𝗻𝗼𝘁 𝗹𝗶𝗸𝗲 𝗽𝗿𝗶𝗰𝗲 𝗱𝗮𝘁𝗮. It is multi-dimensional, requires human judgment in the loop, and changes in ways that need context - not just a number. Getting that onchain, reliably, across 30+ chains, without breaking composability or introducing single points of failure? That took us down some very deep rabbit holes. 𝗪𝗵𝗮𝘁 𝘁𝗵𝗶𝘀 𝗺𝗲𝗮𝗻𝘀 𝗳𝗼𝗿 𝗽𝗿𝗼𝘁𝗼𝗰𝗼𝗹𝘀 𝗶𝘀 𝗿𝗲𝗮𝗹: smart contracts can now read risk. Circuit breakers, dynamic collateral ratios, rating-gated vaults - these become now reality. Building this required deep collaboration with between out engineering teams and partners and providers; weeks of cross-chain routing architecture, and a lot of careful thinking about what "decentralized" actually means. 𝗪𝗲 𝗮𝗿𝗲 𝗻𝗼𝘁 𝗱𝗼𝗻𝗲. More chains to launch on. More capability to unlock. More integrations to build. But today, for the first time, a smart contract can query a live, multi-dimensional risk rating on a tokenized asset and act on it autonomously. That is what we shipped. Particula's Onchain Risk Passport.

52

12 Kommentare

What’s one way you’ve used AI to simplify a task? I use AI to quickly summarize a document, an email, or a graph. Then, ideate by asking: What is missing? What does a concept outline look like? Sometimes I create interactive artifacts, such as a travel to-do list. See this example: https://lnkd.in/ekkxus6B I also use Claude code to draft a POC and to explore technologies. Some of the code can be found on my GitHub. Additionally, I create AI Illustrations for all kinds of documents. Not to forget the dialogues that help me to learn and write some essays. And what is your AI tidbit?

4

4 Kommentare

There is a huge hype around AI all around. And yes, it is useful. And I believe: no, we do not know whether there will be something like general AI based on LLMs or even reliable agents. I really like what Devin (which is one of the more advanced AI tools) recommends: “When to Use Devin TLDR: Treat Devin like a junior engineer. Assign Devin tasks a junior engineer or intern could figure out if provided with sufficient, clear instructions.” https://lnkd.in/e4SD3CRw That's IMO realistic and even if improvements in near future will come for sure, I doubt LLMs will get to “senior level“ anytime soon.

2

Why LangGraph Might Finally Make Agent Workflows Practical I've been hesitant about agentic workflows for a while. Not the idea—multi-agent coordination, persistent memory, human-in-the-loop checkpoints. That all makes sense. But actually pulling it off, in a way that wouldn't collapse under complexity, felt... premature. LangGraph might have changed that. I spent a weekend prototyping with it. What stood out first: it doesn’t try to be magic. No black-box orchestration. You explicitly define the control flow. You see the edges. You can plug in memory, quality gates, even rollback logic—without duct tape. Need a human approval step before an agent submits a pull request? There's a node for that. Need robust state recovery if a task fails mid-stream? Covered. It's like someone finally gave agents a seat at the grown-up table of production software. We're now considering LangGraph for an internal tool that stitches together codegen, test generation, and release approvals. If it holds up, it’ll be the first agentic system we put into production at this scale. Has anyone else tried LangGraph under load or beyond prototyping? Curious about your experience.

7

1 Kommentar

🧠 𝗪𝗛𝗢 𝗪𝗜𝗟𝗟 𝗣𝗥𝗢𝗧𝗘𝗖𝗧 𝗧𝗛𝗘 𝗡𝗘𝗫𝗧 𝗚𝗘𝗡 𝗢𝗙 𝗖𝗢𝗡𝗧𝗘𝗡𝗧 𝗖𝗥𝗘𝗔𝗧𝗢𝗥𝗦? While AI search summaries promise speed, they’re also stealing traffic and value from content creators. After the Cloudflare CEO’s warning, the real question is now: who stands up for authors, publishers and independents? • AI-generated zero-click results Google’s AI Overviews and other bots have driven 56–69% of all searches to zero-clicks, cutting organic traffic by up to 64% • Broken value exchange AI crawlers ingest content without routing users back, breaking the traditional model that fueled the web’s economics. • Economic threat to creators Semrush forecasts show AI search traffic surpassing traditional search by early 2028 if not optimized, and creators will lose even more grip on discovery • Regulation and countermeasures underway Cloudflare rolled out “AI crawler blocking” and “Pay‑Per‑Crawl,” letting publishers control access and monetize their content. 🌍 Who’s already taking action? 💵 Google is piloting micropayments for publishers. 🖱️ OpenAI / Perplexity with low-referral models, <1% outbound clicks . 🤖 Cloudflare is blocking bots by default and enabling Pay‑Per‑Crawl. 🔑 Strategic moves for creators: Use noindex or robots.txt for unwanted crawlers Launch controlled APIs to monetize usage rather than share raw content Optimize for AI search by structuring content for inclusion Advocate for fair compensation frameworks and platform transparency AI isn’t end of creators, t’s a pivot point. But it needs infrastructure, business models and policy adapt. How is your content strategy evolving in the age of zero-click search? — #AI #ContentMarketing #Search #SEO #ContentCreators #DigitalStrategy #Cloudflare #Google #ChatGPT #Perplexity Sources: SEMrush, Google, Cloudflare, ChatGPT/Perplexity

3

2 Kommentare

𝐖𝐡𝐲 𝐘𝐨𝐮 𝐒𝐡𝐨𝐮𝐥𝐝 𝐅𝐨𝐜𝐮𝐬 𝐨𝐧 𝐒𝐞𝐦𝐚𝐧𝐭𝐢𝐜 𝐒𝐞𝐚𝐫𝐜𝐡 𝐈𝐧𝐬𝐭𝐞𝐚𝐝 𝐨𝐟 𝐑𝐞𝐠𝐮𝐥𝐚𝐫 𝐒𝐞𝐚𝐫𝐜𝐡, 𝐚𝐧𝐝 𝐖𝐡𝐲 𝐘𝐨𝐮 𝐒𝐡𝐨𝐮𝐥𝐝 𝐔𝐩𝐠𝐫𝐚𝐝𝐞 𝐘𝐨𝐮𝐫 𝐒𝐄𝐎 𝐒𝐭𝐫𝐚𝐭𝐞𝐠𝐲 𝐟𝐨𝐫 𝐒𝐞𝐦𝐚𝐧𝐭𝐢𝐜 𝐒𝐞𝐚𝐫𝐜𝐡 𝐄𝐧𝐠𝐢𝐧𝐞𝐬 𝐢𝐧 𝟐𝟎𝟐𝟓? Difference Between Regular Search & Semantic Search - 𝐑𝐞𝐠𝐮𝐥𝐚𝐫 𝐒𝐞𝐚𝐫𝐜𝐡 (𝐓𝐫𝐚𝐝𝐢𝐭𝐢𝐨𝐧𝐚𝐥/𝐊𝐞𝐲𝐰𝐨𝐫𝐝-𝐁𝐚𝐬𝐞𝐝 𝐒𝐞𝐚𝐫𝐜𝐡): + Exact words or phrases you type, like finding “dog” in a book’s title or text. + Uses string matching to find documents with your exact words (or close matches like “dogs” or synonyms like “puppy” + Relies on rules or algorithms like BM25 to rank results based on how often and where words appear. + Example: Searching “best running shoes” shows only pages with those exact words or very close matches (e.g., “shoe”). 𝐒𝐞𝐦𝐚𝐧𝐭𝐢𝐜 𝐒𝐞𝐚𝐫𝐜𝐡 (𝐌𝐨𝐝𝐞𝐫𝐧 𝐒𝐞𝐚𝐫𝐜𝐡): + Meaning or intent behind your search, not just the words you type. + Uses artificial intelligence (AI) to connect related ideas and concepts, even if your exact words aren’t in the document. + Searching “pet that barks” shows pages about “dogs” or “canines,” even if “dog” isn’t mentioned, because it gets the idea. If someone want to understand these concepts more in-depth then you should watch the Podcast of Pavel Klimakov on Odys Global

3