Rayyan Ahmed (@Rayyan9477) header
Rayyan Ahmed (@Rayyan9477) profile picture

AI Engineer | Agentic Engineering · AI Agents RAG · MCP · AI System Design · Healthcare AI - LLM Inference - AI RCM 400-days and counting GitHub streak

Joined September 2024
Photo by Rayyan Ahmed (@Rayyan9477)Photo by Rayyan Ahmed (@Rayyan9477)Photo by Rayyan Ahmed (@Rayyan9477)Photo by Rayyan Ahmed (@Rayyan9477)
Hot take: the AGI conversation is making people stupid about healthcare AI. Nobody is getting an “artificial general doctor.” Not this decade. Not in a real health system. You’ll get 20 narrow tools. Imaging. Triage. Coding. Sepsis. Documentation. Drug design. Each one owned by a different vendor. Each one approved for one job. Each one terrifying the moment it starts freelancing. That’s not a failure of ambition. That’s how high-stakes work works. A model that can do everything is a model you cannot validate. If you cannot validate it, you cannot insure it. If you cannot insure it, it does not enter the building. ANI (Narrow Intelligence) is not the consolation prize. It’s the only product that survives contact with reality. Finance, legal, insurance, manufacturing. Same pattern. Specialists print money. Generalists print slides.
🤖 Made with AI
2
40
Institutions keep a durable record about you. You keep nothing comparable about them. Provenance is the other side of that ledger, your open obligations held as versioned beliefs over immutable evidence. A bill arrives 4 months after you cancelled? It doesn't summarise it. It admits the bytes as evidence, types them as a *claim* rather than a fact, detects the contradiction, and reopens the closed case in one serializable transaction. Gemini agents read the evidence. They hold no write credential, enforced by database grant, checked by a linter. A deterministic Kernel is the only thing that writes canonical state. Live on Cloud Run, no login: provenance-web-vaq74wztva-uk… #AllThingsAgenticHackathon @devpost @GeminiApp @Google
1
29
Frontier AI agents still fail most real terminal work. Terminal-Bench 3.0 (the new harder successor to the saturating 2.1 version) just dropped: 74 rigorously reviewed, high-quality tasks across 7 domains of actual valuable computer work. Best models top out around 34–43% resolution rate. Previous versions were hitting ~80%+. This is exactly how we should gauge LLMs and agents, not on toy problems or saturated leaderboards, but on long-horizon, verifiable CLI tasks in realistic environments that expose real failure modes (grounding, recovery, multi-step planning, domain expertise). Great deep-dive from @SnorkelAI on why they fail: snorkel.ai/blog/why-frontier… The gap is still wide. Progress will be measured here.
36
🔥 AI agents just leveled up MCP (the open standard that lets AI use tools & apps) just went fully stateless. No more sessions. Scales like the web. Safer. Can pause mid-task and ask you questions. Claude is already shipping it. This is the USB-C moment for AI.
1
1
27
Claude Opus 5 delivers major gains in agentic coding, computer use, and long-horizon work. While scoring higher on alignment than any previous Claude and remaining under every RSP red line.
1
1
180
After evaluating GPT-5.6 in new ChatGPT desktop app, on our ChatGPT Business plan, the progress is notable. Beyond strong reasoning, the plugin ecosystem enables high-quality automations and deep integrations with external applications. Features like Computer Use, combined with multi-agent orchestration, allow for sophisticated end-to-end workflows that significantly reduce manual work in professional environments. A clear step forward in practical AI tooling. @OpenAI @ChatGPTapp @thsottiaux @gdb
43
The next evolution of agentic AI isn’t bigger single models. it’s smarter orchestration across them.
“Model orchestration is in many ways the natural outgrowth of agentic engineering” Huge thanks to @AndrewYNg and @DeepLearningAI for the deep dive into Sakana AI’s Fugu and Fugu-Ultra. The article highlights how dynamic orchestration allows us to achieve near SOTA performance on benchmarks like GPQA-Diamond, LiveCodeBench Pro, and SWE-Bench Pro without being dependent on a single provider. Full article: deeplearning.ai/the-batch/fu…
1
18
2008 3rd rocket explodes. Company on life support. Everyone screamed “quit pack up.” Yesterday, SpaceX IPO’d at $2.13 TRILLION The largest in history. "When something is important enough, you do it even if the odds are not in your favor." Elon Musk Delusion is just vision the world hasn’t caught up to yet. Stay obsessed. Stay delusional. Grit always wins.
1
41
Interesting
Holy shit it can answer the strawberry question
78
Claude 5 Fable is actually insane. Not just another benchmark flex. it's built for the long, complex, agentic stuff where models usually fall apart. Token efficiency + sustained focus over millions of tokens + real engineering wins (Stripe turning months into days)?
Claude 5 Fable Benchmarks! Holy moly, significant jump even to Mythos
1
72
Agree But Chinese companies will catch up and would provide AI inference at fraction of the cost. Apple would be dependent on others for AI.
Cold take on what comes next: - OpenAI will flourish - Anthropic will continue to be profitable - Google will not catch up to Anthropic or OpenAI - no chinese company will catch up to Anthropic or OpenAI - the highest tier of intelligence will become a luxury product that only companies and multi-millionaires/billionaires can afford - most of the companies that invested massively in them will have massive returns - SpaceX’s AI will be fine and on par with Google by end of year - Nvidia will become the first 10T company
35
AI benchmark wars are starting to feel like Formula 1 for tech companies, everyone flexing numbers, everyone claiming dominance. But beneath the marketing, the progress is actually insane. Models are becoming genuinely useful for coding, research, automation, and real-world work
Introducing Claude Opus 4.8: it builds on Opus 4.7 with sharper judgment, more honesty about its own progress, and the ability to work independently for longer than its predecessors. Available today at the same price.
1
46
Leveling up how we fine-tune AI models 🚀 Shared some practical insights on LLM fine-tuning, trade-offs, and what actually works in real-world setups. If you're building with AI. 👇 Full breakdown: linkedin.com/posts/rayyan-ah… Curious to hear how others are approaching this 👀
1
4
45
Congratulations @karpathy @AnthropicAI , few people shape the direction of AI like this. Excited to see what comes next at the frontier.
Personal update: I've joined Anthropic. I think the next few years at the frontier of LLMs will be especially formative. I am very excited to join the team here and get back to R&D. I remain deeply passionate about education and plan to resume my work on it in time.
1
45
We made it 🏆 3rd prize Track Two: Qwen-Image-2512-Lora-Advertisement. Massive shoutout to the Qwen team for shipping a base model this powerful, and @Ali_TongyiLab @Alibaba_Qwen for an unforgettable competition. One month of training, plenty of doubt, all worth it.
After one month of intense training and another month of rigorous judging, the Qwen-Image LoRA Training Competition has officially concluded! 🏆 We’ve been absolutely floored by the level of technical mastery displayed across every single entry. From the delicate preservation of cultural heritage to the artistic magic of color grading and style transfer, you’ve pushed Qwen-Image to its absolute limits. Massive congratulations to our winners. Check your DM and get in touch with us within one month. First Prize (iPhone 17 Pro Max): @ajie_run_CL @triaakkatuki Second Prize (PS5 Pro): @svntax @liang_zoey2364 @ovi054 @Playmaker_Tech Third Prize ($800 Value Shopping Card): @webdevdeep @ovi054 @dx8152 @galvinlim2001 @Rayyan9477 A huge thank you to everyone who participated. Keep building, keep training!
1
81
My name on this list is something I'll remember forever 🏆 3rd prize, Track Two: AI For Production Qwen-Image-2512-Lora-Advertisement. The talent across every team was insane. Thank you @ModelScope2022 for an incredible competition 🙏 Builders, let's keep going 🚀
After one month of intense training and another month of rigorous judging, the Qwen-Image LoRA Training Competition has officially concluded! 🏆 We’ve been absolutely floored by the level of technical mastery displayed across every single entry. From the delicate preservation of cultural heritage to the artistic magic of color grading and style transfer, you’ve pushed Qwen-Image to its absolute limits. Massive congratulations to our winners. Check your DM and get in touch with us within one month. First Prize (iPhone 17 Pro Max): @ajie_run_CL @triaakkatuki Second Prize (PS5 Pro): @svntax @liang_zoey2364 @ovi054 @Playmaker_Tech Third Prize ($800 Value Shopping Card): @webdevdeep @ovi054 @dx8152 @galvinlim2001 @Rayyan9477 A huge thank you to everyone who participated. Keep building, keep training! We'll be sharing their wonderful work later this month — stay tuned! 🎨
2
71
Surreal moment 🤯 The goal was simple: democratize commercial-grade ad photography so anyone with a single GPU can create studio-quality assets. Honored to see it featured by @Ali_TongyiLab as a 3rd prize winner. Try it, break it, build on it 🙏🚀
📸 Elevate your brand with Pro-Level Ad Photography. Presenting our 3rd prize winner: Qwen-Image-2512-Lora-Advertisement by @Rayyan9477. Trained on 12k+ curated images across 40 advertising categories—all on a single consumer GPU, this 1.7GB LoRA adapter transforms standard photos into commercial-grade advertisement assets with ease. 📍 Try it now: modelscope.ai/models/rayyan9…
2
25
Looks wonderful @deepseek_ai
疑似 DeepSeek V4 技术报告的 benchmark泄露! 图中 的“DeepSeek-V4 技术报告”基准测试(Benchmark)数据,为我们揭示了当前顶尖 AI 大模型的最新竞争格局。从这份涵盖通用能力、推理数学、代码以及智能体(Agent)四大维度的成绩单来看,DeepSeek-V4 展现出了极强的统治力,而大模型的第一梯队也正在经历重新洗牌。 🏆 竞争格局:当前大模型梯队划分 从图表数据中,我们可以清晰地看到五款参评模型的实力分层: 👑 领跑者:DeepSeek-V4在所有 12 项严苛测试中均斩获最高分(State-of-the-Art),实现了跨维度的全面压制。 🥈 最强追赶者:Gemini 3.1 Pro Preview在多项核心指标上紧咬榜首,并在绝大多数测试成绩中超越了 GPT-5.3,是目前这组对比中最具竞争力的对手。 🥉 第三梯队:GPT-5.3 与 Claude Opus 4.6两者互有胜负,依然保持着极高的水准,但在最顶尖的竞争中已稍显疲态。 📏 基准线参照:GPT-4.1作为较早期模型的代表,其数据在这组对比中全面垫底,但也直观地印证了新一代模型技术跨越的幅度之大。 🔍 四大核心能力深度拆解 🧠 1. 综合常识与学科能力 (General) 核心数据:在极具挑战性的 MMLU-Pro 测试中,DeepSeek-V4 (91.2) 和 Gemini 3.1 Pro (90.0) 是唯二突破 90 分大关的模型。 行业洞察:跨学科的专家级知识问答对头部模型已不再是难题。GPT-5.3 (88.4) 和 Claude (86.7) 在这方面稍显落后,知识密度的竞争正在向 90+ 的极限逼近。 🧮 2. 数学与复杂推理 (Reasoning & Math) 核心数据:顶尖数学竞赛基准 AIME 2025 呈现极度“内卷”的态势(DeepSeek-V4 96.4,Gemini 95.0,GPT-5.3 94.6)。 行业洞察:数学是 AI 进步最神速的领域。90+ 的得分意味着这些模型在解决人类高难度奥数题时已经具备了压倒性的优势,各家在这个领域的差距往往只在几道题之间。 💻 3. 编程与工程能力 (Code) 核心数据:在 Codeforces(算法竞赛平台)上,DeepSeek-V4 飙升至 2767 分,拉开显著差距;但在评估修复真实软件工程 Bug 的 SWE-bench Verified 中,所有模型均未突破 60%(最高为 DeepSeek-V4 的 59.6%)。 行业洞察:“写算法题容易,改人类代码难”。模型在纯逻辑生成上已经达到竞赛级选手水平,但在理解和修改复杂的现实商业代码库时,依然存在明显的短板。 🤖 4. 智能体自主行动 (Agent) 核心数据:在模拟网页浏览和执行任务的 WebArena 测试中,全场最高分(DeepSeek-V4)仅为 58.7,GPT-4.1 甚至低至 44.8。 行业洞察:这是全表绝对得分最低的板块。它反映了当前的行业痛点:大模型“做题”和“写文章”能力极强,但如果让它像人类一样自主操作浏览器、跨应用处理多步骤现实任务,成功率依然堪忧。 💡 核心总结这份基准测试不仅是 DeepSeek-V4 强悍实力的“肌肉秀”,也侧面印证了 Gemini 3.1 Pro Preview 在当前技术路线上的极强竞争力。更重要的是,它为行业指明了下一步的攻坚方向——当模型的知识储备和做题能力逼近人类极限时,突破“智能体自主执行 (Agentic tasks)”的现实应用瓶颈,将是决定下一代 AI 霸权的关键。 仅针对图中数据解读,真实情况还有待验证!
1
80
Looks Good, Can't wait get my hands on it. On a lighter note the way @AnthropicAI is building hype is unique and relatively new, but looks like it works. Let's hope it cope up to our expectations.
Introducing Project Glasswing: an urgent initiative to help secure the world’s most critical software. It’s powered by our newest frontier model, Claude Mythos Preview, which can find software vulnerabilities better than all but the most skilled humans. anthropic.com/glasswing
1
66