@LightField Every fix we built makes MCP behave more like an API.
Which raises the question: why not just start there?
Longer writeup on LinkedIn: linkedin.com/feed/update/ur…
English
Ves Stoyanov
157 posts
Ves Stoyanov
@vesko_st
Head of AI at @magicaltome. Ex-Language Researcher at @FacebookAI. Large LMs and multilingual NLP. @JHUCLSP and @Cornell alumn. https://t.co/WTSCasqDI6
Menlo Park, CA Katılım Mayıs 2009
546 Takip Edilen2.1K Takipçiler
How we fixed it at @Lightfield:
→ Meta-tool returns schemas on demand (like --help) → Large results write to files on disk, agent gets a summary and processes with code → Lazy connections + two-layer schema cache kill the startup overhead
English
MCP is becoming the standard for AI tool use. It also breaks at scale.
Tool schemas and results eat tens of thousands of tokens. The agent can't process any of it with code — it's all just text stuck in a prompt.
English
The core problem: everything is loaded eagerly.
Every MCP server dumps full JSON schemas into the prompt. Results inject as raw strings. Your agent has code execution and file I/O, but can't use either on MCP data.
English
How do we know Superhuman intelligence is here? Claude code untangled my tangled commit history. This task is not humanly possible! Trust me I've tried and failed many times before :)
English
Shipping our biggest update to @lightfld since launch.
We combined your data, LLMs, and knowledge of how your business works into tools that don't feel like SaaS — they feel like a helpful colleague.
More features coming soon.
Lightfield@lightfld
Now live: Agent code generation and execution. Your agent can now write and execute code inside a secure sandbox with direct access to your CRM data. This gives it the ability to rapidly navigate thousands of records, produce reports and visualizations, and deliver structured, reliable analysis.
English
“Claude Code is having a ChatGPT moment” [HardFork podcast]
It really resonates. I built two complex features over a weekend with Cursor. Our CPO and designers are prototyping, fixing, and pushing code.
We're at an inflection. AI coding is coming everywhere. Legacy Saas is done!
English
Really excited about the newest open model from MBZUAI! It is a great resource for researchers that need access to truly open models.
MBZUAI@mbzuai
Today, we are releasing a new version of K2 (K2-V2), a 360-open LLM built from scratch as a superior base for reasoning adaptation, while still excelling at core LLM capabilities like conversation, knowledge retrieval, and long-context understanding. K2 fills a major gap: highly capable models with no transparency. Instead of releasing only weights, we’re sharing the full training story — dataset recipes, mid-training checkpoints, logs, code, and evaluation tools. That’s 360-open. What’s inside: • 70B dense transformer engineered as a reasoning-enhanced base model • Native 512K context (extendable via RoPE scaling) • Mid-training reasoning phase • Strong tool-use scaffolding What we’re open-sourcing: • 250M+ reasoning traces (math, planning, multi-step logic) • Full pre- & mid-training data compositions • All mid-training checkpoints • Training logs, code, Eval360 Performance: • GPQA-Diamond: 55.1% mid-training → 69.3% after SFT (strongest fully open 70B model) • KK-8 Logic Puzzles: 83% — competitive with DeepSeek-R1 & OpenAI o3-mini-high • ArenaHard V2: 62.1% — close to Qwen3 235B • Outperforms Qwen2.5-72B and approaches Qwen3-235B despite being smaller and fully transparent. 🔗 The Model: bit.ly/3KIYwuo 🔗Technical Report: bit.ly/49V8h2U 🔗Blog: bit.ly/49V7gb6
English
Today we're officially launching Lightfield. Relational Superintelligence (lnkd.in/gkmSWxw6) for your important day-to-day work building your business. Excited to bring this to everyone.
Henri Liriani@hliriani
We're launching Lightfield today. It's a CRM designed for founders going zero to one—shaped in the past year by hundreds of founders who took a bet on us and now use it daily, and a waitlist of 20,000 more. When starting out, it takes countless hours of talking to customers to figure out what works. From day zero, Lightfield builds and updates itself from your unstructured conversations with customers. It becomes your customer memory: helping you execute deals, communicate without missing details, and understand patterns across your business. It's free to try. Would love to hear what you think: lightfield.app
English
We all thought AGI was around the corner; GPT-5 showed otherwise.
But ASI is here: coding, diagnosis, CFA.
Narrow, yet very useful.
@lightfld, we’re building Relationship Superintelligence:
AI that never forgets a conversation, surfaces what matters, and handles the follow-ups.
English
Ves Stoyanov retweetledi
Getting your first 10 customers is often the hardest part of building a startup, especially given how competitive every space is lately.
Cold emails don’t land, referrals dry up, and every pitch feels like starting from zero. In 2025, AI is changing how founders navigate that 0→1 journey.
That’s why I’m hosting Founder Sales in the AI Era at #SFTechWeek — a candid conversation on how to win your first 10 customers in the AI era, what’s changed, and what hasn’t.
I’ll be joined by an incredible group of founders who’ve lived it:
• Kayvon Beykpour (@kayvz) (Periscope → Twitter → @macroscope_ai)
• Andrei Serban (@andrei_serban) (Fuzzbuzz → Rippling → Console)
• Will Lawrence (@will_lawrenceTO) (Meta → Paxos → @greenliteai)
📅 Oct 7 · 12–1:30pm
🔗 RSVP: partiful.com/e/9F35BgMJXTKv…
@Techweek_ @lightfld
English
Building a brand new communication tool and paradigm with @hliriani and @keithpeiris has been a lot of fun!
Henri Liriani@hliriani
Throughout my career I've designed communication tools: Tome (visual), FB Messenger (with friends), Office (work artifacts). Perhaps unsurprisingly, I’ve come to see the CRM the same way. @lightfld is our ideal version of the daily-use sales computer.
English
Congrats to @douwekiela and team! The @ContextualAI platform looks exciting.
Douwe Kiela@douwekiela
Super excited to share something we’ve been working on for a long time: the @ContextualAI platform is now generally available! SOTA performance across the RAG stack, for each individual component as well as end-to-end 🧵👇
English
Ves Stoyanov retweetledi
We're rebooting Tome to be a different company.
@magicaltome is now an AI assistant for breaking into new enterprise accounts.
Here's a bit on the journey we've been on…
English
🤔 Can LLMs truly plan ahead or are they just predicting the next word? I explore this question by testing AI models with the classic shortest path problem. Discover what this reveals about their planning abilities! 🧠 #AI #MachineLearning #LLM
linkedin.com/posts/ves-sto_…
English
@xiamengzhou @yumeng0818 @danqi_chen I love how simple and effective SimPO is. LHF models have now come full circle back to essentially a version of maximum likelihood training. Great job Mengzhou and team, can't wait to use the method in my everyday work! 6/6
English
@xiamengzhou @yumeng0818 @danqi_chen Simple Preference Optimization (SimPO) [Meng, Xia, and Chen, 2024] removes the need for a reference model and simplifies the optimization objective. It’s elegant: increase the per-token average log probability of the preferred response and reduce it for the losing response. 5/6
English
We have come full circle in Learning From Human Feedback (LHF)! I enjoyed reading the SimPO paper by my brilliant former intern @xiamengzhou (along with @yumeng0818 and brilliant long-time collaborator @danqi_chen). The paper is fascinating. 🧵 1/6
English
@xiamengzhou @yumeng0818 @danqi_chen Everyone who has dabbled in AI/ML knows that reinforcement learning algorithms are notoriously difficult to get working effectively. 3/6
English