Amin Dorostanian

First Snowflake, now a16z. My co-founder Omnya El Massad wrote about the Snowflake signal earlier this week. Today a16z makes the same argument from the investor side: getting an LLM to behave like your best analyst is harder than anyone expected. (Links in comments) That's because training a great analyst isn't a documentation project. It's years of sitting next to someone, learning which tables to trust, why the numbers always dip on Monday mornings, which joins silently blow up. That knowledge only lives in analysts' heads, and in the SQL itself. At a client recently, I inherited a 1,000-line query tracing customers from initial engagement through service delivery. Two senior staff had to walk me through it, sharing years of edge cases, exclusions, and why-this-table-not-that-table that existed nowhere else. If you're reading this, you have many of these. Written by someone who left 18 months ago and took the context with them. Your new analyst is either avoiding them, rewriting them, or using them wrong. If humans struggle to recover that context, imagine what an LLM sees.

57

4 commentaren

200,000 pages per day on a single GPU. DeepSeek's new "Optical Compression" AI is a massive leap for document analysis. 🧠 The cost and complexity of processing thousands of documents can be high. DeepSeek built a new architecture for tackling long documents: Optical compression. Instead of slow, token-by-token text processing, it analyzes an image of a document and converts it into a highly efficient set of "vision tokens." This radically cuts the data load for the language model. 🔹 10x Reduction in Data Input: Achieved while maintaining 97% decoding precision. 🔹 Massive Throughput: A single Nvidia A100 GPU can reportedly process over 200,000 pages per day. 🔹 Superior Efficiency: Far more token-efficient than other benchmarked models. Dramatically reduce the cost of injection pipelines and make scaling AI across platforms far more efficient and affordable. 👉 Follow Ultan O. for more

11

One of the biggest challenges for LLMs has been accuracy when handling massive context.  Recursive Language Models (RLMs) -- a powerful new framework from MIT CSAIL that fundamentally rethinks how AI handles context. RLMs treats context as an external resource instead of stuffing everything into a single prompt. What makes it groundbreaking: --Unlimited context capacity – Instead of overwhelming the model with massive prompts, RLM treats the full input as an external “database.” The model writes Python code to query segments, call itself recursively, and assemble insights. --Better performance at scale – On benchmarks like OOLONG and CodeQA, RLMs outperform GPT‑5 and other context agents by 10–25%, even when processing tens of millions of tokens. --Transparent & efficient – The reasoning steps happen via generated code, making the logic inspectable. Why it matters: --Transforms long-context reasoning from a memory limit issue to a software orchestration problem --Enables scalable, explainable AI—critical for applications in legal review, scientific papers, and long-form analytics --Opens the door to general-purpose AI agents that can handle truly enormous bodies of information You can read the full paper here: https://lnkd.in/e8wuwb8a

9

DeepSeek AI may have just solved one of the biggest challenges facing the use of LLMs - Context length. In the paper, DeepSeek describe the use of a new form of OCR based vision LM which is used to produce language output. From what we can see in the results, the model is able to provide 10X the context length with >95% accuracy which is huge. All of this has been demonstrated with a very small model, so very exciting stuff. Really keen to see where the follow on research takes this work. Let me know what your thoughts are on the paper in the comments. #AI #LLM #VLM #SLM #GPT #OCR https://lnkd.in/gmHXG2jh

23

3 commentaren

📈 More LLMs, higher quality!! 💰 Gemini, Cursor (Composer) & GPT-5 reviewed Claude plans & implementation on 27 PRs Only 35% of issues were found by multiple LLMs. The rest of the findings? Completely unique to a single tool. Claude caught a CSRF bypass and WebSocket auth gaps that nobody else touched. GPT-5 flagged flow-breaking type errors. Cursor caught race conditions and a few bugs. Dropping any one tool would leave blind spots. Claude is the strongest reviewer, however, in Claude Code it has the most context as well. Subjectively, Cursor Bugbot delivers best insights & actual bugs on Github - Composer run on local changes is not as productive. My next steps are 1) Quantify quality of Copilot & Bugbot reviews on Github 2) try claude-octopus (link in comments) & acpx for stability of the setup

14

3 commentaren

We just shipped accent detection for voice AI calls at Roark (YC W25) 🎧 This isn’t LLM-based. It’s a dedicated ML model (ECAPA-TDNN) trained on 500k+ labeled speech samples, reaching ~89% accuracy across 15 English accent variants. Most teams building voice agents focus on what is said. Very few have visibility into how it’s said. Accent consistency turns out to matter a lot — especially when you’ve configured a specific persona (e.g. Australian support agent, British concierge). With this, you can: - Detect accent per speaker segment directly from audio - Track accent shifts mid-call - Measure an “accent stability score” across a conversation - See probability distributions (e.g. 45% American, 35% British) One simple use case: Automatically flag calls where your TTS voice drifts from its intended accent. We’re starting with 16 English accents (US, UK, AU, IN, etc.) and expanding from here. Now available as a metric package inside Roark. Feels like a missing piece for teams running voice in production.

116

10 commentaren

Your best prompt engineer is your domain expert. So why build an AI product at all? It's a fair question, and one I hear a lot at the POC stage: "My expert already gets great results from Claude or GPT. They know the business better than any model does. With a bit of training, they'll only get better at prompting. Why build anything on top of that?" The premise is right. A model has read the internet. Your expert has lived your business. A well-prompted expert will often beat an average AI product. And the expert never leaves the room. It's their deliverable. They own the output and have to verify it. That doesn't change, nor should it. So the goal was never to remove the expert. It's to make their work faster and their handoffs cleaner. Think about how work moves across a team. A task passes from one person to the next, across hundreds of cases, and quality drops at every handoff, because no two people share the same instincts. A product captures that expertise once, as instructions, examples, and checks, so the same standard travels with the work. The expert then spends their time verifying good output instead of rebuilding it. A good answer is also not a reliable answer. A skilled prompter is excellent most of the time. But "most of the time" is not a business process. A product adds the plumbing, the evaluations, guardrails, and fallbacks, that the expert can trust and sign off on quickly. And a chat window has no memory, no data, and no audit trail. A product connects securely to your systems, controls who sees what, and records why a decision was made. So the next person isn't starting from zero. So it was never expert versus AI. AI will not replace the expert. It assists the expert, speeds up their work, and carries their standard through every handoff. That is the real value. Not a smarter answer than your expert could give. The same quality, produced faster, and held to the same standard every time it changes hands. If your product can't clear that bar, you don't have a product yet. That's the question to ask at POC stage, before the budget scales. #AI #AITransformation #MachineLearning #AIStrategy #DomainExpertise

7

Here's a project I previously worked on with MIT Professional Education for their Applied Data Science certification program. 𝐁𝐚𝐜𝐤𝐠𝐫𝐨𝐮𝐧𝐝 The EdTech industry has been surging in the past decade, with a compound annual growth rate (CAGR) of 10.26% from 2018 to 2023. Due to the industry's rapid growth and the attraction of new customers, new companies have emerged in the industry. These new companies can reach a wider audience with the availability and ease of use of digital marketing resources. 𝐏𝐫𝐨𝐣𝐞𝐜𝐭 𝐆𝐨𝐚𝐥 ExtraaLearn is an initial-stage startup that offers programs on cutting-edge technologies to students and professionals to help them upskill/reskill. With a large number of leads being generated, one of ExtraaLearn's biggest challenges is to identify which leads are more likely to convert into paying customers, so that the company can allocate resources accordingly. As a data scientist in the company, I need to analyze and build an ML model that can identify leads more likely to convert into paying customers, identify factors driving the lead conversion process, and create a profile of leads more likely to convert. 𝐂𝐨𝐧𝐜𝐥𝐮𝐬𝐢𝐨𝐧𝐬 Based on both the decision tree and random forest models I created, the top three most important features are time spent on the website, first interaction through the website, and medium profile completion. Based on the class 1 recall scores, the best performing model appears to be the tuned decision tree model, as it has a recall score of 0.86 on the testing data. The exploratory data analysis suggests that leads who are most likely to convert into paying customers are people in their 40s or 50s who are either working professionals or unemployed, and have their first interaction with ExtraaLearn through the website. 𝐑𝐞𝐜𝐨𝐦𝐦𝐞𝐧𝐝𝐚𝐭𝐢𝐨𝐧𝐬 We saw that time spent on the website and leads having their first interaction through the website are the most important factors in converting leads into paying customers. In order to convert more leads, ExtraaLearn should prioritize the website. The company should push leads to interact with the website first. In addition, ExtraaLearn should also create incentives for leads to stay on the website for as long as possible. ExtraaLearn should also refocus its marketing efforts towards demographics who are most likely to convert. The demographic that should be prioritized is middle-aged individuals (40s and 50s) who are either working professionals or unemployed. With this marketing shift, ExtraaLearn should see a greater return-on-investment with lead conversions. ⭐ If you like to learn more, please click on this link: https://lnkd.in/drNhpRMg #datascience #python #datavisualization #opentowork #networking

9