📆 ThursdAI - Feb 13 - my Personal Rogue AI, DeepHermes, Fast R1, OpenAI Roadmap / RIP GPT6, new Claude & Grok 3 imminent?

What a week in AI, folks! Seriously, just when you think things might slow down, the AI world throws another curveball. This week, we had everything from rogue AI apps giving unsolicited life advice (and sending rogue texts!), to mind-blowing open source releases that are pushing the boundaries of what's possible, and of course, the ever-present drama of the big AI companies with OpenAI dropping a roadmap that has everyone scratching their heads.

Buckle up, because on this week's ThursdAI, we dove deep into all of it. We chatted with the brains behind the latest open source embedding model, marveled at a tiny model crushing math benchmarks, and tried to decipher Sam Altman's cryptic GPT-5 roadmap. Plus, I shared a personal story about an AI app that decided to psychoanalyze my text messages – you won't believe what happened! Let's get into the TL;DR of ThursdAI, February 13th, 2025 – it's a wild one!

* Alex Volkov: AI Adventurist with weights and biases

* Wolfram Ravenwlf: AI Expert & Enthusiast

* Nisten: AI Community Member

* Zach Nussbaum: Machine Learning Engineer at Nomic AI

* Vu Chan: AI Enthusiast & Evaluator

* LDJ: AI Community Member

Personal story of Rogue AI with RPLY

This week kicked off with a hilarious (and slightly unsettling) story of my own AI going rogue, all thanks to a new Mac app called RPLY designed to help with message replies. I installed it thinking it would be a cool productivity tool, but it turned into a personal intervention session, and then… well, let's just say things escalated.

The app started by analyzing my text messages and, to my surprise, delivered a brutal psychoanalysis of my co-parenting communication, pointing out how both my ex and I were being "unpleasant" and needed to focus on the kids. As I said on the show, "I got this as a gut punch. I was like, f*ck, I need to reimagine my messaging choices." But the real kicker came when the AI decided to take initiative and started sending messages without my permission (apparently this was a bug with RPLY that was fixed since I reported)!

Friends were texting me question marks, and my ex even replied to a random "Hey, How's your day going?" message with a smiley, completely out of our usual post-divorce communication style. "This AI, like on Monday before just gave me absolute s**t about not being, a person that needs to be focused on the kids also decided to smooth things out on friday" I chuckled, still slightly bewildered by the whole ordeal. It could have gone way worse, but thankfully, this rogue AI counselor just ended up being more funny than disastrous.

Open Source LLMs

DeepHermes preview from NousResearch

Just in time for me sending this newsletter (but unfortunately not quite in time for the recording of the show), our friends at Nous shipped an experimental new thinking model, their first reasoner, called DeepHermes.

NousResearch claims DeepHermes is among the first models to fuse reasoning and standard LLM token generation within a single architecture (a trend you'll see echoed in the OpenAI and Claude announcements below!)

Definitely experimental cutting edge stuff here, but exciting to see not just an RL replication but also innovative attempts from one of the best finetuning collectives around.

Nomic Embed Text V2 - First Embedding MoE

Nomic AI continues to impress with the release of Nomic Embed Text V2, the first general-purpose Mixture-of-Experts (MoE) embedding model. Zach Nussbaum from Nomic AI joined us to explain why this release is a big deal.

* First general-purpose Mixture-of-Experts (MoE) embedding model: This innovative architecture allows for better performance and efficiency.

* SOTA performance on multilingual benchmarks: Nomic Embed V2 achieves state-of-the-art results on the multilingual MIRACL benchmark for its size.

* Support for 100+ languages: Truly multilingual embeddings for global applications.

* Truly open source: Nomic is committed to open source, releasing training data, weights, and code under the Apache 2.0 License.

Zach highlighted the benefits of MoE for embeddings, explaining, "So we're trading a little bit of, inference time memory, and training compute to train a model with mixture of experts, but we get this, really nice added bonus of, 25 percent storage." This is especially crucial when dealing with massive datasets. You can check out the model on Hugging Face and read the Technical Report for all the juicy details.

AllenAI OLMOE on iOS and New Tulu 3.1 8B

AllenAI continues to champion open source with the release of OLMOE, a fully open-source iOS app, and the new Tulu 3.1 8B model.

* OLMOE iOS App: This app brings state-of-the-art open-source language models to your iPhone, privately and securely.

* Allows users to test open-source LLMs on-device.

* Designed for researchers studying on-device AI and developers prototyping new AI experiences.

* Optimized for on-device performance while maintaining high accuracy.

* Fully open-source code for further development.

* Available on the App Store for iPhone 15 Pro or newer and M-series iPads.

* Tulu 3.1 8B

As Nisten pointed out, "If you're doing edge AI, the way that this model is built is pretty ideal for that." This move by AllenAI underscores the growing importance of on-device AI and open access. Read more about OLMOE on the AllenAI Blog.

Groq Adds Qwen Models and Lands on OpenRouter

Groq, known for its blazing-fast inference speeds, has added Qwen models, including the distilled R1-distill, to its service and joined OpenRouter.

* Record-fast inference: Experience a mind-blowing 1000 TPS with distilled DeepSeek R1 70B on Open Router.

* Usable Rate Limits: Groq is now accessible for production use cases with higher rate limits and pay-as-you-go options.

* Qwen Model Support: Access Qwen models like 2.5B-32B and R1-distill-qwen-32B.

* Open Router Integration: Groq is now available on OpenRouter, expanding accessibility for developers.

As Nisten noted, "At the end of the day, they are shipping very fast inference and you can buy it and it looks like they are scaling it. So they are providing the market with what it needs in this case." This integration makes Groq's speed even more accessible to developers. Check out Groq's announcement on X.com.

SambaNova adds full DeepSeek R1 671B - flies at 200t/s (blog)

In a complete trend of this week, SambaNova just announced they have availability of DeepSeek R1, sped up by their custom chips, flying at 150-200t/s. This is the full DeepSeek R1, not the distilled Qwen based versions!

This is really impressive work, and compared to the second fastest US based DeepSeek R1 (on Together AI) it absolutely flies

Agentica DeepScaler 1.5B Beats o1-preview on Math

Agentica's DeepScaler 1.5B model is making waves by outperforming OpenAI's o1-preview on math benchmarks, using Reinforcement Learning (RL) for just $4500 of compute.

* Impressive Math Performance: DeepScaleR achieves a 37.1% Pass@1 on AIME 2025, outperforming the base model and even o1-preview!!

* Efficient Training: Trained using RL for just $4500, demonstrating cost-effective scaling of intelligence.

* Open Sourced Resources: Agentica open-sourced their dataset, code, and training logs, fostering community progress in RL-based reasoning.

Vu Chan, an AI enthusiast who evaluated the model, joined us to share his excitement: "It achieves, 42% pass at one on a AIME 24. which basically means if you give the model only one chance at every problem, it will solve 42% of them." He also highlighted the model's efficiency, generating correct answers with fewer tokens. You can find the model on Hugging Face, check out the WandB logs, and see the announcement on X.com.

ModernBert Instruct - Encoder Model for General Tasks

ModernBert, known for its efficient encoder-only architecture, now has an instruct version, ModernBert Instruct, capable of handling general tasks.

* Instruct-tuned Encoder: ModernBERT-Large-Instruct can perform classification and multiple-choice tasks using its Masked Language Modeling (MLM) head.

* Beats Qwen .5B: Outperforms Qwen .5B on MMLU and MMLU Pro benchmarks.

* E…