Yamini Ananth

Synthetic Sciences (YC W26) is building AI co-scientists for end-to-end scientific research. Scientific research is still a marathon of fragmented work across papers, code, compute, and writing. Each iteration forces researchers to rebuild context, rerun experiments, and stitch results back into drafts. You can’t parallelize yourself, so progress bottlenecks on attention and context. Synthetic Sciences lets scientists delegate the full research loop to AI co-scientists: literature review and synthesis grounded in your project, hypothesis generation, experiment planning, code execution in containerized environments, GPU runs, monitoring, analysis, and publication-ready drafts (LaTeX, figures, slides). They’ve battle-tested synsci most on ML research so far and are rapidly expanding across domains. Biology mode is already SOTA on BixBench Verified (92%). They’re also launching Flywheel Mode, which helps teams turn production traces and data into custom, private, task-specific models they can own and improve over time. Thanks to strong open-weight models and advances in post-training and RL, these models often end up faster, smarter, and cheaper than frontier model APIs for their specific workflows. Try it or book a demo: syntheticsciences.ai Congrats on the launch, Ishaan Gangwani and Aayam Bansal! https://lnkd.in/gU_u7kAq

433

45 Comments

🎮 Four LLMs. One Mario. Who actually moves? At UC Berkeley’s Hao AI Lab, researchers ran a little experiment: Make leading AI models play Super Mario Bros. No hand-holding. Just frame-by-frame gameplay. The goal? See who can adapt, decide, and not die instantly. 🏆 Claude 3.7 made it past the first pipe. 💀 GPT-4o, Claude 3.5, and Gemini 1.5 Pro? Still pondering the meaning of “run”. This isn’t about gaming. It’s about real-world intelligence. Because if your AI can’t handle a Goomba… good luck trusting it with your GTM strategy. ⚠️ Flashy scores and benchmark charts are cute. But what really matters? ✔️ On-the-fly decisions ✔️ Real-time reasoning ✔️ Tactical clarity under pressure Heidi gets that. She doesn’t play Mario—but she does help you dominate meetings like a boss. No rage quits. Just real outcomes. 👇 Seen a funnier AI fail? Drop it in the comments. And if you want an AI that actually gets sh*t done, Join the waitlist (and go grab our productivity playbooks) : https://hub.hi-heidi.ai/ #generativeAI #supermario #llm #Claude3 #GPT4o #Gemini #AIleadership #smartagents #realtimethinking #heidiAI #aitools

2

Anysphere, maker of AI coding assistant Cursor, has reportedly raised $900M at a $9B valuation, led by Thrive Capital, with participation from a16z and Accel, according to the Financial Times. Previously, Cursor raised $100M at a $2.6B valuation. The startup has now raised over $173M. Investors like Index Ventures and Benchmark are eager to join, but existing backers are doubling down. Rival AI coding startup Windsurf is also seeking funding at a $3B valuation. #AI #CursorAI #Anysphere #AIStartups #TechFunding #DealonAI

1

Anysphere is the company behind AI coding assistant Cursor Says an IPO is nowhere in sight, even after hitting $1B annualized revenue and raising $2.3B at a $29.3B valuation. CEO Michael T. says the focus now is on deeper AI automation, custom LLMs, and building Cursor into a full end-to-end development engine capable of tackling complex bugs and multi-step engineering workflows. Cursor is also shifting hard into enterprise mode with spend controls, billing tools, code-review automation, and upcoming agentic features designed to handle “tasks that take humans weeks.” Key Points: • IPO off the table; all energy on product expansion. • Cursor now using in-house LLMs that generate “more code than almost any other model”. • Pricing moved to usage-based as dev teams rely on Cursor for hours, not minutes. • New enterprise tools: cost management, billing groups, full-cycle code reviews. • Next phase: agentic systems that fix hard bugs end-to-end. Will Cursor become the first true AI engineer that replaces not just coding but entire dev workflows? Follow MediaBlizz for more AI, tech, and innovation deep dives. #Cursor #Anysphere #AICoding #LLMs #TechNews #AIStartups #EnterpriseAI #SoftwareEngineering #MediaBlizz

3

When LLMs Should Not Be Used in Production Large Language Models are amazing — but they’re not a fit for every use case. If your system demands high accuracy, real-time speed, sensitive data handling, or edge-case reliability, blindly deploying LLMs can create bigger problems than they solve. I wrote about when to think twice before putting LLMs into production and how to use them safely. 👉 Read the full article here: https://lnkd.in/ghVqb_HP Have you ever tried using an LLM in production and hit unexpected issues? #technicalscripter2026 #day20 #llms #aiproducts #generativeai #machinelearning #ai #developerexperience #geeksforgeeks #systemdesign

10

Small LLMs are bad at multi-step math. Everyone knows this. The usual fix — sample 40 chains, majority vote — works but burns tokens. What if you could catch a bad reasoning step *as it happens* instead of after the fact? That's the question we explored for last semester for our CS685 at UMass Amherst. The result: Dual-Track CoT, a two-model loop where a Decomposer proposes steps and an Evaluator scores them in real time, with a hard token budget forcing the system to be efficient. Some things we learned the hard way: → Fine-tuning both models for their specific roles is non-negotiable → The Evaluator is the bottleneck — when it's wrong, it's worse than having no Evaluator at all → Symbolic matching beats learned embeddings for anything involving numbers Full blog post linked below — would love to hear what you think. https://lnkd.in/eRmUJVu6 #NLP #LLM #MachineLearning #UMassAmherst #Research #CS685

18

With all the hype around Large Language models, I decided to look under the hood. How they work and what powers LLMs. Through this exploratory journey, I wrote a post about components of LLM's, inspired heavily by 3b1b's series on deep learning: https://lnkd.in/gPVgfVKy

58

Just wrapped up a deep read of NaturalThoughts: Selecting and Distilling Reasoning Traces for General Reasoning Tasks from FAIR at Meta. It’s one of those papers that doesn’t just report numbers - it invites you to rethink how we approach reasoning in LLMs. What makes this work stand out is its clarity of focus: not all reasoning traces are created equal. The authors move beyond the traditional "scale everything" mindset and ask better questions. Which types of reasoning examples actually make models better at general reasoning? What kinds of traces help them think - not just produce answers with lots of tokens? NaturalThoughts is built on the insight that performance gains come not just from more data, but from smarter data curation. The team shows that selecting examples based on difficulty - like when multiple teacher models disagree - or based on diversity in reasoning strategies leads to stronger students. In some cases, these strategies outperformed popular distillation datasets like OpenThoughts and LIMO across STEM reasoning benchmarks. They also explore mixing System-1 and System-2 style reasoning traces, allowing student models to be both more efficient and more adaptable at test time. That part struck me, because it feels closer to how humans actually reason - we switch between quick intuition and deliberate analysis depending on context. If you're working on instruction tuning, alignment, or data-centric model improvement, this is worth your attention. It’s not just about finding answers, it's about learning what kinds of reasoning help models grow. #AIResearch #ReasoningInAI #InstructionTuning #ModelInterpretability #DataCentricAI #FutureOfAI #LinkedIn #connections #AI #ML #DL #NLP #GenerativeAI #PromptEngineering paper link : https://lnkd.in/e85qQ2KJ https://lnkd.in/enu6EF2R

22

4 Comments