# 📆 ThursdAI - Sep 4 - Codex Rises, Anthropic Raises $13B, Nous plays poker, Apple speeds up VLMs &...

Source: https://www.youtube.com/watch?v=ltPx0QMH_Rg
Recap page: https://rapidrecap.app/video/ltPx0QMH_Rg
Generated: 2025-09-05T01:01:42.607+00:00

---
## Quick Overview

The ThursdAI September 4th episode covered significant AI advancements, including Codex's improved performance with GPT-5, Anthropic's massive $13 billion funding round, the release of Nous Research's Hermes 4 models and Husky Holdem benchmark, Apple's speedy VLMs, and OpenAI's GPT real-time updates.

**Key Points:**
- Codex, powered by GPT-5, demonstrates significantly improved performance, with usage reportedly increasing 10x in two weeks, making it a solid coding assistant option.
- Anthropic secured a substantial $13 billion in Series F funding, valuing the company at $183 billion post-money, indicating strong investor confidence and significant growth.
- Nous Research released Hermes 4 models (405B and 70B fine-tuned from Llama 3.1, and 14B fine-tuned from Qwen 3), focusing on neutral alignment and demonstrating strong performance on non-traditional benchmarks like EQBench and their new Husky Holdem poker bot evaluation.
- Apple introduced Fast VLM 7B, a vision encoder noted for its speed, being 85 times faster to first token compared to peers, and is gaining traction on Hugging Face.
- OpenAI launched GPT real-time to General Availability, offering improvements in speech-to-text models for lower latency and better instruction following, though still not matching text-based LLM intelligence.
- Google released Embedding Gemma, an open-source embedding model with 300 million parameters, ranking first on the MTB multilingual benchmark and suitable for on-device RAG applications.
- The open-source landscape saw the release of Tensent's Hunan MT (a 7B machine translation model) and Switzerland's Uppercas 8B and 70B multilingual LLMs trained on 15 trillion tokens across 1,800 languages, both released under the Apache 2.0 license.

**Context:** This episode of ThursdAI, hosted by Alex Fov, features co-hosts Yan Pelleg and LDJ, with guests Roger Jin and Besh Bves from Nous Research, and Quindla Kramer discussing AI developments. The discussion covers a wide range of topics from coding agents and LLM performance to significant funding rounds and new open-source releases, reflecting the rapid pace of innovation in the AI field.

## Detailed Analysis

The ThursdAI September 4th episode highlighted several key developments in the AI landscape. Codex, now integrated with GPT-5, has shown remarkable improvement, leading to a tenfold increase in usage over two weeks, positioning it as a strong contender for coding tasks. In the realm of large-scale funding, Anthropic raised an impressive $13 billion in Series F, reaching a $183 billion valuation. Nous Research contributed to the open-source community with the release of their Hermes 4 models, which are fine-tuned versions of Llama 3.1 and Qwen 3, emphasizing neutral alignment and showing promise on diverse benchmarks. They also introduced the Husky Holdem benchmark, a novel evaluation for LLMs to create poker bots. Apple entered the VLM space with Fast VLM 7B, a notably fast vision encoder. OpenAI made its GPT real-time speech-to-text model generally available, improving latency and instruction following capabilities, though it still trails text-based LLMs in raw intelligence. Google also released Embedding Gemma, a compact and efficient open-source embedding model that tops the MTB multilingual benchmark. The open-source segment also featured Tensent's Hunan MT, a specialized translation model, and Switzerland's Uppercas LLMs, notable for their multilingual capabilities and open-source nature. The discussion touched upon the strategic importance of these releases for both enterprise and individual developers, emphasizing the growing ecosystem of AI tools and models.

### News & Announcements

- Codex performance boost with GPT-5
- Anthropic raises $13B
- Apple's Fast VLM 7B
- OpenAI GPT real-time GA
- Google Embedding Gemma release

### Open Source Releases

- Nous Research's Hermes 4 models (Llama 3.1 & Qwen 3 fine-tunes)
- Tensent Hunan MT (7B translation model)
- Switzerland's Uppercas 8B & 70B (1800 languages)
- Apache 2.0 license highlighted for Uppercas

### Benchmarks & Evaluations

- Nous Research's Husky Holdem (LLM-generated poker bots)
- EQBench performance for Hermes 4
- MTB multilingual benchmark for Embedding Gemma

### Model Performance & Capabilities

- Codex's 10x usage increase
- Hermes 4 shows generalization beyond math/code
- GPT real-time improves speech-to-text latency
- Embedding Gemma's efficiency (300M parameters, 200MB footprint)

### Company & Funding News

- Anthropic's $13B Series F funding at $183B valuation
- OpenAI's free user project updates
- OpenAI hires VPs from Statsig

### Guest Insights (Nous Research)

- Roger Jin and Besh Bves discuss Hermes 4 data mix (5M samples, 56B tokens) and training cutoff based on 'soul' preservation
- Besh explains Husky Holdem benchmark as LLMs coding poker bots, not playing directly

### Guest Insights (Voice AI)

- Quindla Kramer discusses OpenAI's GPT real-time (GA), its limitations in matching text LLM intelligence, and the trend towards multi-agent systems in voice AI and coding agents