Dmitry Yemelyanov

Our agents found GPT-5 team's educational background. These are the schools building the future. Our analysis of GPT-5's 14 core architects reveals interesting patterns about where breakthrough AI innovation originates. MIT actually dominates with 4 connections, followed by Carnegie Mellon and University of Warsaw, then Stanford and École Polytechnique/EPFL with 2 each. But here's the interesting part: The international presence is massive. → University of Warsaw (Poland) - 2 members → École Polytechnique/EPFL (Switzerland) - 2 members → University of Waterloo (Canada) - 1 member → VU University Amsterdam (Netherlands) - 1 member → Sciences Po Paris (France) - 1 member The team combines: → Theoretical CS from CMU → Physics PhDs from Harvard → Mathematics olympiad winners from Poland → Engineering scholars from Switzerland Nearly every member has major competition wins as well. The takeaway is that AI teams aren't built by poaching from the same 3 schools. The minds that built GPT-5 came from 14 different universities across multiple countries. If you're only recruiting from Stanford and MIT, you're missing half the talent pool. The future belongs to teams that think differently, not teams that went to the same places. This is just one team analysis. Our agents can research any company’s educational, advisor networks, contact details, hiring strategies, etc. Reach out to see what we can uncover for you.

123

8 Comments

"With only 27 million parameters, HRM achieves exceptional performance on complex reasoning tasks using only 1000 training samples. The model operates without pre-training or CoT data, yet achieves nearly perfect performance on challenging tasks including complex Sudoku puzzles and optimal path finding in large mazes. Furthermore, HRM outperforms much larger models with significantly longer context windows on the Abstraction and Reasoning Corpus (ARC)" https://lnkd.in/gZU-AkKp https://lnkd.in/g9dqj_WA

173

19 Comments

GPT-5 is the best coding model in the world and is now the default model in Cursor! Cursor: “It’s the most intelligent coding model our team has tested.” There is an amazing story about how this model became so much better at coding, featuring some innovative core research + behind-the-scenes efforts by the startups solutions team. In building GPT-5 the research team invested heavily in coding improvements and trained a model that showed large gains on the leading coding and SWE benchmarks. TL;DR: it was SOTA in coding. However, as we tested early alpha snapshots of the model GPT-5 with several of our most sophisticated coding customers, we realized there were still some gaps in performance and personality that only showed up when you used the models in products like Cursor. Making the experience great is about more than just the raw intelligence, which was solid - it’s also about how the model actually behaves in the product that matters for creating magical end-user experiences. For example, the first model snapshot would silently make many sequential tool calls without summarizing to the user what it’s doing or why, which created a less helpful developer experience. Or, it would produce front-end designs that weren’t as beautiful and clean, even though they were functional and produced with fewer errors than prior models. Seeing these gaps, a small group of startups solutions team members and engineers sprung into action. They decided to try to tackle this problem and bring the customer voice more squarely into our research process, working backwards from the actual tasks that startups needed the model to perform in their product. The key realization is that it’s not enough to saturate the public benchmarks; we had to make sure the models work and feel right in products like Cursor. A small team sprinted to work backward from these customer use cases and create representative evals, graders, and rubrics for scoring the model’s improvements on the key dimensions that matter - some of which were subjective by nature, like front-end design. In close collaboration with the research team, these inputs were incorporated into the training of subsequent model snapshots. The cross-functional team ran rapid cycles of development/testing/iteration to steer the behavior and performance of the model. The result is an *exceptional* coding model. It not only excels at core coding tasks - producing high-quality code, fixing bugs, editing code, and completing repo-level refactors - but also has a collaborative personality and an ability to generate strikingly clean & beautiful front-end designs. It’s been an incredibly exciting time at OpenAI. We are loving GPT-5 and I’m excited for developers everywhere to experience this model! https://openai.com/gpt-5/

284

15 Comments

Water boils at 100°C. Not at 98°C. That final push? It changes everything. Product development is no different. Idea validation — At 98°C, you assume users will need it. — At 100°C, you validate with real users and iterate from their feedback. (Think Airbnb listing their own apartment before raising funds.) MVP build — At 98°C, you rush to release something minimal. — At 100°C, you still move fast, but polish enough for someone to use it daily without friction. User experience — At 98°C, the app works, but feels clunky. — At 100°C, small design details turn first-time users into long-term users. (Think of Notion’s obsessive focus on UX before it took off.) Go-to-market — At 98°C, you launch and wait. — At 100°C, you show up, talk to users, fix what’s broken, repeat. Tech & team — At 98°C, you hire help. — At 100°C, you partner with people who think like co-builders and care about your vision. The difference between a good product and a great product often comes down to that extra 2 degrees of effort. Always go all in

42

16 Comments

The 4-month timeline that's killing your voice AI projects: Every SaaS company thinks they can add voice AI in a few weeks. The reality? Even well-funded companies with full engineering teams need 4-5 months to get enterprise-level voice AI right. Just came out of a meeting with an enterprise company handling 20,000 calls per month. Their CTO: "We want voice AI live by next quarter. How hard can it be?" Me: "What's your current call classification accuracy?" Him: "Classification? We just need the AI to answer phones." This is where 90% of voice AI projects fail. Companies think it's about the voice. It's actually about: → Domain expertise - Understanding conversation design patterns → Testing infrastructure - Real-world scenarios, not demo scripts → Telephony integration - Call routing, transfers, fallbacks → Observability systems - Knowing when and why calls fail → Speech technology depth - Handling accents, interruptions, noise The companies that nail it in 4 weeks? They're building basic chatbots, not enterprise voice systems. The ones that take 5+ months? They're building agents that can handle a stressed customer calling about payroll issues while their toddler screams in the background. Big difference. If you're a SaaS company planning voice AI integration, budget for months, not weeks. And find someone who's done this before. Because the "how hard can it be?" approach has a 100% failure rate. P.S - The SaaS company? We're starting with call classification and working up from there. Smart move.

61

18 Comments

Prior to OpenAI's GPT-5 release, Anthropic's Sonnet model was the undisputed king of AI coding. But it turns out that GPT-5 is significantly better than Sonnet in one key way: recovering from mistakes. Today we're releasing Recovery-Bench: a new benchmark for measuring how well AI agents can recover from errors and corrupted states. AI coding tools like Claude Code are changing the way software is written - at Letta , everyone on the team is a heavy user of either Claude Code or Cursor (or both), and the majority of our code written is "AI-assisted". But coding agents today have a fatal flaw: they often get confused by past mistakes and accumulate errors over time, making it difficult to engage in long coding sessions with the same agent. Instead, most engineers resort to clearing the agents memory and starting fresh. This is where Recovery-Bench comes in - it's a systematic way to measure a model's ability to recover from past mistakes. What are the main takeaways? 1. The best performing models overall are not the same as the best "recovering" models. Claude Sonnet 4 leads the pack in overall coding ability compared to GPT-5 in 4th place, but on recovery bench, GPT-5 dominates. This means that GPT-5 is more likely to be able to "recover" from a past mistake during a coding session and less likely to start "looping", compared to Claude Sonnet. This result also implies that if your Sonnet-backed coding agent fails during a session, you might want to swap out the model for GPT-5 to help you get unstuck. If you're using a model-agnostic agent platform like Letta, this is as simple as flipping a switch. 2. Recovery from failed states is a challenging, unsolved task. New techniques in context management will be required to solve the recovery problem and continual learning for LLMs. For more on Recovery-Bench, check out our blog post and GitHub repo below. At Letta, we are building self-improving, perpetual agents that can learn from experience and adapt over time. If you’re excited by our mission and want to contribute towards work like Recovery-Bench, apply to join our team.

143

15 Comments

Google's new LLM can run on just 0.5 GB RAM. (let's fine-tune it 100% locally) Google released Gemma 3 270M, a new model for hyper-efficient local AI! We'll fine-tune this model and make it very smart at playing chess and predict the next move. Tech stack: - Unsloth AI for efficient fine-tuning. - Hugging Face transformers to run it locally. Here's how you can do it in 5 easy steps: 1️⃣ Load the model ↳ We start by loading the Gemma 3 270M and its tokenizer using Unsloth. 2️⃣ Define LoRA config ↳ We'll use LoRA for efficient fine-tuning. ↳ To do this, we use Unsloth's PEFT and specify: - The model - LoRA low-rank (r) - Layers for fine-tuning (target_modules) 3️⃣ Load dataset ↳ We'll fine-tune Gemma 3 to make it extremly smart at playing chess. ↳ Given a set of previous move (one move missing) & the final result it has to predict the missing move. ↳ In order to do this we're using the ChessInstruct dataset from HuggingFace. 4️⃣ Prepare dataset ↳ Next, we use a conversation style dataset to fine-tune Gemma 3. ↳ The standardize_data_formats method converts the dataset to the correct format for finetuning purposes! 5️⃣ Define Trainer and train the model ↳ We create a Trainer object by specifying the training config, like learning rate, model, tokenizer, and more, and initiate training. Done! This gives us the fine-tuned Gemma 3 that can find the exact missing chess move instead of randomly generating some moves. Find the complete fine-tuning guide by Unsloth in the comments. _____ Share this with your network if you found this insightful ♻️ Follow me (Akshay Pachaar) for more insights and tutorials on AI and Machine Learning!

1,488

67 Comments

🟢 Another surprising breakthrough in AI—this one few expected so soon. A big deal in AI just happened which has not got ( so far) much coverage. OpenAI has just announced that a general-purpose reasoning model achieved gold medal-level performance at the 2025 International Mathematical Olympiad (IMO)—solving 5 out of 6 problems, under the same time limits as humans, and without any external tools. That’s not just impressive—it's unprecedented. Previous AI wins (Go, Dota, Poker) came from systems narrowly engineered for a single task. This wasn’t that. As OpenAI explains: “This isn’t an IMO-specific model. It’s a reasoning LLM that incorporates new experimental general-purpose techniques… IMO problems were the perfect challenge: proofs are pages long and hard to verify. This model thinks for hours—not seconds—and it’s more efficient with that thinking.” ➡️ To put this in context: ✅ In early 2024, AI models were tackling grade-school math (e.g. GSM8K). ✅ Then came high school-level (MATH dataset), followed by AIME. ✅ Now, gold-level performance at IMO, globally regarded as the ultimate pre-university math challenge. 🧠 What’s different? This model reasons deeply and deliberately—something AI has historically struggled with. As articles from Simon Willison, The Decoder, and MathArena explain, this result far outpaces models like Gemini 2.5 Pro and DeepSeek-R1, which didn’t reach even bronze-level scores on the same test. The MathArena benchmark underscores why this is a leap: the problems demand novel reasoning and multi-step proofs, not just pattern recognition or retrieval. We're now seeing models that can persist, reason, reflect—and win. 💡 Big takeaway: AI still fails at many things. But progress at the frontier is accelerating—especially on the hardest tasks. Structured reasoning under time pressure. Multi-step deduction. Agentic problem solving. It’s happening. This result was led by a small team at OpenAI, headed by Alexander Wei. It wasn’t expected—even within OpenAI. It shows how fast general-purpose models are improving when guided by the right innovations. Want to see what the model actually produced? Its solutions are available ⬇️ ⬇️ We may be approaching a new era—where AI doesn’t just assist in math and science, but contributes meaningfully to discovery. 🔗 Results: https://lnkd.in/gse5rNW9 🔗 Book2: https://lnkd.in/eJfZ3PWt 🔗 Book1: https://bit.ly/4b31PEG 🔗 Alignment paper: https://lnkd.in/ez8cV-jK #AI #OpenAI #Mathematics #MachineLearning #MathOlympiad #AgenticAI #ScientificDiscovery #Technology #FutureOfWork

26