Uploaded March 2026 | Updated September 2026, 3 weeks ago
Mistral is one of the world's leading frontier model labs, and has just raised $900m to build their European data center hub. Last year marked their first ventures into multimodal models, with Pixtral and then Voxtral in 2025, ending the year with a monster €1.7B fundraise at a €10B premoney (mistral.ai/news/mistral-ai-raises-1-7-b-to-accelerate-technological-progress-with-ai) - the largest European AI fundraise ever.
This week they launched Voxtral TTS (mistral.ai/news/voxtral-tts) and it is filling in yet another important gap in the open source frontier for all modalities of industrially used AI models. We were excited to catch up with Pavan, Voxtral lead, and Guillaume, Chief Scientist!
00:00 Welcome and Guests
00:22 Announcing Voxtral TTS
01:41 Architecture and Codec
02:53 Understanding vs Generation
04:50 Flow Matching Explained
07:25 Real Time Voice Agents
13:40 Efficiency and Model Strategy
14:53 Voice Agents Vision
17:56 Enterprise Deployment and Forge
23:39 Fine Tuning and Personalization
25:22 Enterprise Voice Personalization
26:09 Long Form Speech Models
26:58 Real Time Encoder Advances
27:45 Scaling Context for TTS
28:53 What Makes Small Models
30:23 Merging Capabilities Strategy
32:04 Voice Meets Video Latency
33:05 Open Source Mission
35:51 Lean and Formal Proofs
38:40 Reasoning Transfer and Agents
40:25 Next Training Frontiers
42:20 Hiring and Global Teams
43:05 AI for Science Partnerships
44:19 Forward Deployed Engineering
47:01 Real World Evals and Wrap Up
Mistral is one of the world's leading frontier model labs, and has just raised $900m to build their European data center hub. Last year marked their first ventures into multimodal models, with Pixtral and then Voxtral in 2025, ending the year with a monster €1.7B fundraise at a €10B premoney (mistral.ai/news/mistral-ai-raises-1-7-b-to-accelerate-technological-progress-with-ai) - the largest European AI fundraise ever.
This week they launched Voxtral TTS (mistral.ai/news/voxtral-tts) and it is filling in yet another important gap in the open source frontier for all modalities of industrially used AI models. We were excited to catch up with Pavan, Voxtral lead, and Guillaume, Chief Scientist!
00:00 Welcome and Guests
00:22 Announcing Voxtral TTS
01:41 Architecture and Codec
02:53 Understanding vs Generation
04:50 Flow Matching Explained
07:25 Real Time Voice Agents
13:40 Efficiency and Model Strategy
14:53 Voice Agents Vision
17:56 Enterprise Deployment and Forge
23:39 Fine Tuning and Personalization
25:22 Enterprise Voice Personalization
26:09 Long Form Speech Models
26:58 Real Time Encoder Advances
27:45 Scaling Context for TTS
28:53 What Makes Small Models
30:23 Merging Capabilities Strategy
32:04 Voice Meets Video Latency
33:05 Open Source Mission
35:51 Lean and Formal Proofs
38:40 Reasoning Transfer and Agents
40:25 Next Training Frontiers
42:20 Hiring and Global Teams
43:05 AI for Science Partnerships
44:19 Forward Deployed Engineering
47:01 Real World Evals and Wrap Up
![[State of AI Papers 2025] Fixing Research with Social Signals, OCR & Implementation — Team AlphaXiv](https://i.ytimg.com/vi/T0mZJjl_dsA/mqdefault.jpg "[State of AI Papers 2025] Fixing Research with Social Signals, OCR & Implementation — Team AlphaXiv
From late-night dorm-room hacking sessions at Stanford to building the most-used platform for navigating AI research, the founders of *AlphaXiv* have spent the last few years watching the archive firehose explode from thousands to 30,000+ papers per month—and building the intelligent layer on top that researchers, applied engineers, and even VCs now rely on to make sense of it all. We caught up with *Raj, Rayhahn, and the AlphaXiv team* live at *NeurIPS 2025* to dig into the origin story (spoiler: it started as a class project with a view comment button next to archive paragraphs), why they beat Hugging Face Papers by obsessing over UI and getting Lauras authors to comment directly on their own work, how theyve evolved from comments to feeds to benchmarks to AI assistants that answer paper questions using Geminis 1M context window, whats broken in the academic review process (20% AI-generated reviews at ICLR, quality collapsing under exponential submission growth), why *papers are becoming less important* than implementations, Docker containers, and real-world usability, their vision for making it trivial to spin up any trending papers code directly in your browser, and the power law they see every day: archive has 2.4M papers, but applied researchers only care about the top 0.1% that are actually implementable—and AlphaXiv is building the ranking, tooling, and social signals to surface exactly those.
We discuss:
* The *origin story:* from a late-night web dev class project (view comment buttons on archive PDFs) to going viral on LinkedIn and becoming a full-time startup backed by Stanford advisors
* Why AlphaXiv beat *Hugging Face Papers:* better UI, direct commenting on the paper itself, and early traction with high-signal authors like the Laura/DPO/Llama teams
* The evolution: *comments → feed → benchmarks → AI assistant,* using social signal (views, comments, Twitter) to filter the archive firehose and surface what actually matters
* How they handle *PDF parsing and OCR:* DeepSeek Coder for cost/accuracy, Mistral OCR APIs, and passing diagrams directly into multimodal models like Claude
* Why *semantic search alone fails:* 3M papers, buzzword overload, and the need to weight by social signal (views, engagement) to return actually relevant results
* The *state of academic publishing:* exponential submission growth (30k/month in CS), 20% AI-generated reviews at ICLR, review quality collapsing, and the rise of AI-written papers flooding the system
* Why *papers are becoming obsolete:* the future is implementations, Docker containers, and interactive sandboxes—not PDFs describing what code does
* The *power law:* 2.4M papers on archive, but applied researchers (Spotify, Expedia, Nintendo) only care about the top 0.1% that are actually implementable and useful
* AlphaXivs roadmap: *Docker containers for papers,* agent-automated setup, ranking by implementation ease, and making it trivial to spin up trending work directly in your browser
* Papers of the year: *Tiny Recursive Models (TRMs),* evolutionary strategies at hyperscale, Agent R1 (RL-trained agents for complex reasoning), and the rise of AI for science (Agent Laboratory, Bionemo)
* The *Qwen vs. DeepSeek* dynamic: Qwens breakout year, DeepSeeks continued influence (OCR models triggering a flurry of competitive releases), and the open-source fine-tuning ecosystem around both
* Why *AI reviewers* could help (Stanfords Agent Reviewer, using AI as a linter for clarity and quality) but assessing novelty and lit review remains hard without search + social signal
* The vision: AlphaXiv as the *tool layer for research*—not a social network, but the best place to discover, understand, and implement the ideas that actually matter in AI
—
AlphaXiv Team
* AlphaXiv: https://alphaxiv.org
* X: https://x.com/alphaxiv
Where to find Latent Space
* X: https://x.com/latentspacepod
* Substack: https://www.latent.space/
00:00:00 Introduction: The Origin Story of AlphaXiv
00:03:12 Beating Hugging Face Papers: What Made AlphaXiv Different
00:04:17 From Comments to Discovery: Building the Research Feed
00:04:55 PDF Parsing and OCR: The Technical Infrastructure
00:06:01 Becoming a Company: Beyond Papers to Research Artifacts
00:06:37 Benchmarks and State of the Art: Replacing Papers with Code
00:07:54 Personalization and Search: Building the Best Research Assistant
00:08:54 Docker Containers and Reproducibility: Making Papers Runnable
00:10:52 Favorite Papers: Tiny Recursive Models and Evolutionary Strategies
00:13:15 AI for Science: Agent Laboratory and the Future of Research Automation
00:18:41 Agent R1 and RL for Long-Horizon Tasks
00:20:54 Qwens Year and the Open Source Landscape
00:22:43 The Crisis in Academic Publishing: AI Slop and Review Quality
00:25:04 Search and Social Signals: Why Semantic Search Isnt Enough
00:28:15 The Future of Research: Beyond the PDF
00:31:17 Implementation Ease as the New Ranking Signal")
