ThursdAI - Apr 3rd - OpenAI Goes Open?! Gemini Crushes Math, AI Actors Go Hollywood & MCP, Now with Observability?

Woo! Welcome back to ThursdAI, show number 99! Can you believe it? We are one show away from hitting the big 100, which is just wild to me. And speaking of milestones, we just crossed 100,000 downloads on Substack alone! [Insert celebratory sound effect here 🎉]. Honestly, knowing so many of you tune in every week genuinely fills me with joy, but also a real commitment to keep bringing you the the high-signal, zero-fluff AI news you count on. Thank you for being part of this amazing community! 🙏

And what a week it's been! I started out busy at work, playing with the native image generation in ChatGPT like everyone else (all 130 million of us!), and then I looked at my notes for today… an absolute mountain of updates. Seriously, one of those weeks where open source just exploded, big companies dropped major news, and the vision/video space is producing stuff that's crossing the uncanny valley.

We’ve got OpenAI teasing a big open source release (yes, OpenAI might actually be open again!), Gemini 2.5 showing superhuman math skills, Amazon stepping into the agent ring, truly mind-blowing AI character generation from Meta, and a personal update on making the Model Context Protocol (MCP) observable. Plus, we had some fantastic guests join us live!

So buckle up, grab your coffee (or whatever gets you through the AI whirlwind), because we have a lot to cover. Let's dive in! (as always, show notes and links in the end)

OpenAI Makes Waves: Open Source Tease, Tough Evals & Billions Raised

It feels like OpenAI was determined to dominate the headlines this week, hitting us from multiple angles.

First, the potentially massive news: OpenAI is planning to release a new open source model in the "coming months"! Kevin Weil tweeted that they're working on a "highly capable open language model" and are actively seeking developer feedback through dedicated sessions (sign up here if interested) to "get this right." Word on the street is that this could be a powerful reasoning model. Sam Altman also cheekily added they won't slap on a Llama-style <700M user license limit. Seeing OpenAI potentially re-embrace its "Open" roots with a potentially SOTA model is huge. We'll be watching like hawks!

Second, they dropped PaperBench, a brutal new benchmark evaluating an AI's ability to replicate ICML 2024 research papers from scratch (read paper, write code, run experiments, match results - no peeking at original code!). It's incredibly detailed (>8,300 tasks) and even includes meta-evaluation for the LLM judge they built (Nano-Eval framework also open sourced). The kicker? Claude 3.5 Sonnet (New) came out on top with just 21.0% replication score (human PhDs got 41.4%). Props to OpenAI for releasing an eval where they don’t even win. That’s what real benchmarking integrity looks like. You can find the code on GitHub and read the full paper here.

Third, the casual 40 Billion Dollarsfunding round led by SoftBank. Valuing the company at 300 Billion. Yes, Billion with a B. More than Coke, more than Disney. The blog post was hilariously short for such a massive number. They also mentioned500 million weekly ChatGPT usersand the insane onboarding rate (1M users/hr!) thanks to native image generation, especially seeing huge growth in India. The scale is just mind-boggling.

Oh, and for fun, try the new grumpy, EMO "Monday" voice in advanced voice mode. It's surprisingly entertaining.

Open Source Powerhouses: Nomic & OpenHands Deliver SOTA

Beyond the OpenAI buzz, the open source community delivered some absolute gems, and we had guests from two key projects join us!

Nomic Embed Multimodal: SOTA Embeddings for Visual Docs

Our friends at Nomic AI are back with a killer release! We had Zach Nussbaum on the show discussing Nomic Embed Multimodal. These are new 3B & 7B parameter embedding models (available on Hugging Face) built on Alibaba's excellent Qwen2.5-VL. They achieved SOTA on visual document retrieval by cleverly embedding interleaved text-image sequences – perfect for PDFs and complex webpages.

Zach highlighted that they chose the Qwen base because high-performing open VLMs under 3B params are still scarce, making it a solid foundation. Importantly, the 7B model comes with an Apache 2.0 license, and they've open sourced weights, code, and data. They offer both a powerful multi-vector version (ColNomic) and a faster single-vector one. Huge congrats to Nomic!

OpenHands LM 32B & Agent: Accessible SOTA Coding

Remember OpenDevin? It evolved into OpenHands, and the team just dropped their own OpenHands LM 32B! We chatted with co-founder Xingyao "Elle" Wang about this impressive Qwen 2.5 finetune (MIT licensed, on Hugging Face).

It hits a remarkable 37.2% on SWE-Bench Verified (a coding benchmark measuring real-world repo tasks), competing with much larger models. Elle stressed they didn't just chase code completion scores; they focused on tuning for agentic capabilities – tool use, planning, self-correction – using trajectories from their contamination-free Switch Execution dataset. This focus seems to be paying off, as the OpenHands agent also snagged the #2 spot on the brand new Live SWE-Bench leaderboard! Plus, the 32B model runs locally on a single 3090, making this power accessible. You can also try their managed OpenHands Cloud ($50 free credit). Amazing progress from this team!

Frontiers: Diffusion LMs & Superhuman Math

Two other developments pushed the boundaries this week:

Dream 7B: A Diffusion Language Model Challenger?

This one's fascinating conceptually. Researchers unveiled Dream 7B, a language model based on diffusion, not auto-regression. The benchmarks they shared show it competing strongly with top 7-8B models, and absolutely crushing tasks like Sudoku (81% vs <50% for others), potentially due to its parallel processing nature being better for global constraints. It's an exciting hint at alternative architectures, but the model weights aren't out yet, so we can't verify or play with it. Still, one to watch!

Gemini 2.5 Obliterates Olympiad Math (24.4% on USAMO!)

We already knew Gemini 2.5 was good, but wow. New results dropped showing its performance on the USA Math Olympiad (USAMO) – problems so hard most top models score under 5%. Gemini 2.5 Pro scored an incredible 24.4%!

The gap between it and everything else is massive, highlighting the power of its reasoning and thinking capabilities (which you can inspect via its traces!). Having used it for complex tasks myself (like wrestling with tax forms!), I can attest to its depth. It's free in the Gemini app – go try it!

Agents, Compute & Making MCP Observable

Amazon's Nova Act Agent & The Need for Access

Amazon entered the agent chat with Nova Act, designed for web browser actions. They claim it beats Claude 3.5 and OpenAI's QA model on some benchmarks, possibly leveraging acquired Adept talent. But... it's only available via an SDK with a request form. As Yam rightly pointed out on the show, these agent claims mean little until we can actually use them in the real world!

CoreWeave + NVIDIA = Insane Speeds

Hardware keeps accelerating. CoreWeave announced hitting 800 Tokens/sec on Llama 3.1 405B using NVIDIA's new GB200 Blackwell chips, and 33,000 T/s on Llama 2 70B with H200s. Inference is getting fast.

This Week's Buzz: Let's Make MCP Observable!

Okay, my personal mission this week builds on the growing Model Context Protocol (MCP) momentum. MCP is potentially the "HTTP for agents," enabling tool interoperability. But as tool use moves external, we lose visibility, making debugging and security harder.

That's why I'm launching the Observable Tools initiative. The goal: integrate observability into the MCP standard itself. Right now, that link redirects to a GitHub discussion where I've proposed using the OpenTelemetry (OTel) standard to add tracing to MCP interactions. This would give developers clear visibility into tool usage, regardless of their observability platform.

I need your help! Please check out the proposal, join the discussion, and show your…