Mohamed Elgendy

YellowDog.ai just set a 10x benchmark uplift in scale computing, delivering 40,000 tasks per second (TPS) and managing 100,000 compute nodes in the cloud. What's even more interesting, that's 2x IBM Symphony and opens an intriguing pathway for these until-know closed/captive systems.

11

In October 2021, I argued on the Hugging Face blog that the mega-model arms race was a dead end. The model that prompted the post was Megatron-Turing NLG 530B. Today, Qwen3-14B scores higher on harder benchmarks and runs on a consumer GPU. It has 38× fewer parameters. The prediction was right. The mechanism was engineering: Flash Attention, quantization, llama.cpp, MoE, Unsloth AI. Smaller, cheaper, open models won. But the prediction missed something. By December 2025, Chinese-origin models accounted for 63% of all new derivatives on Hugging Face. Qwen overtook Meta Llama in cumulative downloads. DeepSeek AI R1 matched OpenAI o1 under an MIT license. The base layer of the global open ecosystem shifted to Hangzhou. This happened on merit. Better multilingual support. More permissive licenses. Faster iteration. Then this week happened. Yesterday: Arcee AI shipped Trinity Large Thinking — a 400B reasoning model trained from scratch in the U.S., Apache 2.0. Today: Google released Gemma 4 under Apache 2.0, the first time the Gemma family ships fully permissive. NVIDIA's Nemotron 3 Super landed two weeks ago. OpenAI released GPT-OSS last summer. My new piece, "Open Source Won, But Changed Owners", traces the full history of open models from BERT to today, with two charts you'll love to steal for your talks 😉 The base-layer fight is on: who builds the model that the world will fine-tune? Link in comments.

37

6 Comments

Benchmarks say frontier models (GPT, Opus, Gemini Pro, etc) are roughly comparable. My agent just proved otherwise. We've been building Minerva, a platform that uses long-running AI agents to explore complex scientific problems. My agent takes about 20 minutes to run and makes repeated calls to the underlying model, with access to a variety of tools it chooses autonomously. The agent plans its own strategy for solving the problem - I don't tell it how to do it, just what I need it to do and the tools it has access to. I've been testing it using supposedly competitive models - the same architecture, same tools, same task, but a different model under the hood. The results weren't comparable - not even close. The first gap was tool-calling behaviour. One model made basic, predictable tool calls. Another worked out how to use sophisticated parameters, things like regex filters, to extract far more nuanced results from the same tools. Same API, completely different level of competence. The second gap was strategy. The exact same agent, given the exact same problem, adopted a radically different approach depending on which model was driving it. The planning and sequencing weren't just better or worse. They were fundamentally different. When I got an LLM to compare the final outputs, it described the best result as being "in a completely different league from everything else". This matters because most people choosing a model start with the leaderboards. They see a few percentage points separating the frontier models and treat them as interchangeable. In agentic work, they might not be. Because we're making so many calls to the LLM, a subtle difference in earlier calls can get compounded as the agent continues. There's a cost dimension too. The winning agent run cost me around $25 in tokens to answer a single question. When you're iterating on agent design, that adds up fast. But consider what $25 actually bought - output that could fundamentally reshape a scientist's understanding of a problem they thought was intractable. That could push research in a direction they'd never have considered. In that context it's legitimate to think that $25 is actually quite cheap. I think Minerva is heading somewhere interesting. My lesson: you can't read, benchmark or theorise your way to understanding this stuff - you have to build things and run them. Practical experience is everything.

15

2 Comments

Are “dead” dashboards a good sign for your business? A recent guest on my podcast makes the argument that it can be, especially if your framing is that the purpose of data is to change the business gradually. In Lucas’s own words: “As I approach this, I try to put myself into the shoes of the business user — the head of sales, or an account manager, whoever it might be. When they see something, I want them to have a reaction of, ‘Okay, now I want to do this.’ If they don’t have a reaction that causes them to take an action, then I haven’t really done a good job of putting data in front of them. I think that’s where this adoption challenge and repeat-usage challenge comes in, where dashboards die pretty quickly. You set it up — and then I had a person come up to me at a conference and ask me about this. They said, ‘Look, I just built this dashboard and we spent three months on it, and within a couple of months, nobody’s using it anymore. What the heck, I feel like I failed at my job.' I told them, ‘No, you did not.’ The purpose of data, the purpose of insights and analytics, is to change the business gradually. As you build really great reports, they cause business users to ask follow-up questions and to evolve the business and slowly change it, continuing to optimize it into something new. When the next questions are different from the ones three months ago, that actually means you did a great job, because you moved the conversation. You helped the organization evolve. So the fact that your dashboards die is actually a great thing. With Orion, I started talking about sentient dashboards, where it’s never really a static thing anymore. It continues to evolve with you. As Orion sees questions arise in different functions, it starts to modify the analysis it does. As it sees multiple people asking similar questions, its proactive approach also changes in different directions.” – Lucas Thelosen , CEO @ Gravity 👇 Excerpt from episode 181 of Experiencing Data, one of the top 2% of all podcasts in the world. 🔈Listen here: https://bit.ly/47ufC6u

6

2 Comments

Most teams underestimate how hard structured extraction really is — until they try to map a real document into a real JSON schema. At Hirize.AI, we learned early that this is where simple OpenAI API calls break immediately on real documents. They hit limits like 100 keys + 5 nested levels, and they fail on multi-page context and hierarchical structure. Traditional API-style extraction breaks on: - Deeply nested schemas - Multi-page context - Conditional logic & Cross-references That’s why we rebuilt our entire structured extraction engine around hierarchy, continuity, and schema fidelity — not “LLM optimism.” Here’s what that looks like in practice: A customer came to us with a 12-page schema instruction manual full of conditionals, exceptions, and domain rules. LLM-based tools couldn’t handle it — they failed at 100 keys, broke hierarchy, or flattened the structure entirely. We parsed the schema directly into our engine, and extracted data from a 120-page document with 97%+ accuracy — preserving hierarchy, references, and structure end-to-end. Why I’m sharing this? Because this is the part of document AI that separates demos from production. Anyone can extract text. Very few can produce clean, validated, structured outputs that plug directly into: - SQL pipelines - EHR systems - Financial models - Operational dashboards This is the difference between “interesting technology” and infrastructure businesses rely on. The more businesses try to scale AI, the clearer this becomes: If the structure is wrong, the entire downstream system breaks. And that’s exactly the layer we’re solving at Hirize.

18

1 Comment

Most teams are choosing LLMs the way they chose cloud providers in 2015. The core truth...No single LLM dominates across all dimensions. Each one leads in a different axis: • Claude → software engineering + safety • Gemini → multimodal + enterprise compliance • GPT-5 / o-series → deep reasoning + math • Grok → real-time social intelligence Choosing correctly is a systems decision, not a popularity contest 1. Claude (Anthropic) Claude leads where correctness matters more than creativity. • 82% SWE-bench Verified, the highest real-world coding score • Lowest jailbreak success rate (4.7% ASR) • Strong long-form coherence and refusal discipline • Designed around Constitutional AI, not post-hoc filters Tradeoffs: No public fine-tuning. Limited multimodal support. Conservative refusals. Best fit: software engineering teams, legal workflows, research summarization, high-risk domains where hallucinations are unacceptable. 2. Gemini (Google) Gemini is built for scale, modality, and regulated environments. • 1M token context window • Native video and long-audio understanding • Strong balanced reasoning (GPQA leader) • FedRAMP High + ISO 42001 certified • Deep Workspace and Vertex AI integration Tradeoffs: Higher hallucination rate than Claude. Heavier enterprise stack. Best fit: multimodal pipelines, enterprise data analysis, government and healthcare deployments. 3. ChatGPT / GPT-5 (OpenAI) GPT dominates abstract reasoning and customization. • 100% AIME math benchmarks • Best ARC-AGI scores among frontier models • Most complete fine-tuning stack (SFT, DPO, RL) • Only provider with a free standalone moderation API Tradeoffs: Higher factual hallucination rate. Reasoning models add latency. Best fit: math, planning, research, agent systems, products that require heavy customization. 4. Grok (xAI) Grok optimizes for immediacy, not caution. • Native real-time X/Twitter access • Fast responses • Strong live signal awareness Tradeoffs: Weakest published safety guarantees. Sparse compliance documentation. Best fit: social intelligence, trend analysis, real-time monitoring. Why benchmarks are no longer enough - HumanEval is effectively solved. - MMLU is saturating above 90%. - Even math benchmarks are hitting ceilings. The real gaps now show up in: • Hallucination behavior • Latency under reasoning load • Compliance and data residency • Fine-tuning availability • Cost of being confidently wrong ARC-AGI v2 makes this clear: reasoning ability still varies massively, even when other scores look identical. The practical rule - Claude minimizes harm - Gemini maximizes coverage - GPT maximizes reasoning - Grok maximizes immediacy Stop asking “Which LLM is best?” Start asking “Which failure mode can my system survive?” That question determines whether your AI product scales or silently breaks. PS:) Like, and SAVE! . . .

182

6 Comments

AI adoption in P&C is not linear and it's definitely not tidy. One pattern keeps popping up across the P&C teams I speak with - most of us are somewhere on a spectrum that loosely shakes out into three levels of capability. 🤖 Level 1: Experimentation This is the curiosity phase - where most of us have been hanging out in 2025. People test tools, play with prompts and see what might stick. ⚠️ The real risk is staying here too long and ending up with scattered tools, siloed insights and too many agents with too little value added. 🤖 🤖 Level 2: Intentional use This is when AI shifts from tinkering to supporting real workflows. It reduces manual lift, speeds up decisions and creates consistency. This is where we start to see real ROI - time/money saved and decisions are made faster with better data and more accurate insights. 🤖 🤖 🤖 Level three: Embedded capability AI becomes part of how teams and organisations operate. It connects to real systems and understands context. It improves decisions without someone needing to remember which tool to use or the right questions to ask. ✨ This level is where ways of working change (for the better I'd argue). My prediction for next year is more P&C teams will finally move from curiosity to real capability, and some organisations will take the first steps toward embedding AI into the way they operate. 🙋‍♀️ Where do you sit? I'm halfway between Level 1 and 2 most days. If you're currently shaping your AI strategy for P&C next year, join us next week for small session focused on this topic and more with Lisa-Marie Moegle (Sitemate), Graham Moody (JobAdder), and Bianca Hill (Decidr) at Updoc's offices in Sydney. 👉 Limited spots left so RSVP here: https://luma.com/xaazou74

14

1 Comment

AIs can make you dumb and AIs can make you smart and it all depends on how you use it. This is how I use LLMs: 1️⃣ As a debating partner to refine my ideas: I share my views, and thesis with LLMs and ask it critique or debate with me. This helps me refine my own ideas and helps me identify the blind spots in my understanding. Even many of my LinkedIn posts are first 2️⃣ As a personalized teacher to help me pick new skills: I ask it to share resources for learning. Then I share my understanding with it to ask it to evaluate. Then I ask it to give me tasks to practice. (as someone who runs a training academy, at the risk of losing clients to AI, I would still say that it is a phenomenal tool for learning) 3️⃣ Rewrite for me: I sense that sarcasm and judgmental tones often seep into my posts. So, I share my write up with LinkedIn and ask it reduce those elements. And also point to me which part of what I wrote sounds too judgmental. 4️⃣ As my personal assistant: Adding comments to my codes or even elaborating more on my core thesis or documenting my models are all a lot of labor work. Wherever possible I leave it to the AI to do it. And I completely rely on AI for adding decorative elements like emojis and icons. What I will not do with AI is just give a topic and ask it to give me a write up. Or just upload a client email and ask it to create a model. How do you use AI?

15

1 Comment