Easy AI Daily | June 23, 2025
A roundup of AI industry news for June 23, 2025, covering model releases, research advances, funding, tools, and hardware.
Model Releases and Updates
- Sakana AI Reinforcement Learning Teachers (RLTs): Trains 7B models via reinforcement learning to generate step-by-step explanations that accelerate Chain-of-Thought learning in larger models — the small model acts as a "teacher" rather than the LLM being trained directly. Details
- Mistral Small 3.2: Improved instruction following and function calling, experimental FP8 quantization; tool calling in GGUF/transformers largely fixed. Update
- Google Magenta RealTime: 800M-parameter open model for real-time instrumental music generation from mixing parameters; runs in free-tier Colab; claimed as the first open real-time music generation model. Project | GitHub | Demo
- Arcee AI AFM-4.5B: 4.5B-parameter foundation model built by extending Llama 3, five months in development, targeting sub-10B scenarios. Blog
- OpenThinker3-7B: New open 7B reasoning model, +33% over DeepSeek-R1-Distill-Qwen-7B on reasoning benchmarks. Release
- STORM text-to-video: Inserts Mamba layers to compress video inputs 8x; scores 70.6% on MVBench, surpassing GPT-4o; combines SigLIP vision encoder with Qwen2-VL. Highlights
- OmniGen2: Open image editing model based on Qwen-VL-2.5 (an open Flux.1 Kontext alternative), with decoupled text/image decoding paths, no flash-attn required; strong single-image editing, good multi-image editing, beating OmniConsistency and BAGEL. Project | GitHub | Hugging Face
- PPO vs GRPO: PPO uses clipped objectives and a value model for stable general-purpose learning; GRPO removes the value model and normalizes rewards within groups, better suited to reasoning-heavy tasks. Analysis
- DINOv2 on ImageNet-1k: Aggressively optimizing k-nn accuracy on ImageNet-1k suggests overfitting evaluation metrics can still yield better generalizing models. Analysis
- Spectral clipping: New optimization technique that caps singular values (e.g., values above β=8 set to 8), unlike Muon which pushes them toward 1; improves stability. Blog
- Torchtune dataset packing: Fixes OOM issues from packing, supports pre-tokenized packed datasets, dynamic packing planned to speed reasoning-model training. PR #2819
- Harvey AI: $300M Series E at a $5B valuation, led by Kleiner Perkins and Coatue, for R&D and market expansion. Announcement
- Replit: ARR surpassed $100M, a 10x increase since end of 2024. Announcement
- ElevenLabs 11ai: Voice-first AI assistant supporting MCP, integrated with Perplexity, Linear, and Slack. Product
- MCP Validator: New release supports the 2025-06-18 MCP spec, adding OAuth 2.1 auth, structured tool output, and GitHub Actions compliance templates. GitHub
- Unsloth: Fixed Gemma 3 8-bit/4-bit quantization RuntimeError; users should upgrade unsloth-zoo and unsloth, with PyTorch 2.7cu12.6 recommended.
- Cursor: New pricing caused rate-limit confusion; some users report Gemini 2.5 Pro looping/verbosity issues and suggest switching to Sonnet 4. Pricing
- Chisel CLI: Profiles AMD MI300X via cloud servers at $1.99/hour, auto-syncing code and running rocprof. GitHub
- Notebook LM: Now indexes and retrieves images (charts, figures) from PDFs.
- Yuval Noah Harari compared the AI revolution to billions of "digital immigrants" arriving visa-free at light speed, disrupting jobs and power structures; he called for attention to governance challenges. Talk
- Mechanize trains AI agents in simulated environments for professional skills (engineering, law), aiming for a "fully automated economy." Intro
- AI code quality debate: Developers report inexperienced teams producing low-quality AI code that inflates maintenance needs and demand for senior engineers, though tools like Claude Code may improve this. Discussion
- Blackwell B200 / RTX 5090: Both run Gemma 3 stably; the B200 nearly exhausts VRAM training a 27B model; 5090 pricing in Europe dropped to about €2,200.
- AMD Ryzen AI Max 395: With 128GB LPDDR5x, runs 70B+ models in LM Studio at 3–4 tokens/second; memory allocation needs driver improvements. Demo
- Neutrino: Open-source eBPF-based GPU kernel-level profiling tool with assembly-level probing and memory access visualization, accepted at USENIX OSDI '25. GitHub
- Arch-Agent-7B: 7B model scores 69.85 on multi-step agentic workflows, surpassing GPT-4.1 (68.89); strong at function calling and complex task planning. Model card
- Sherlog-MCP: IPython-based shared workspace enabling multi-agent collaboration with persistent results, sidestepping context window limits. GitHub
Research and Algorithms
Industry and Funding
Tools and Platforms
Tutorial: Getting Started with n8n
An 8.5-hour n8n video tutorial is recommended; the first 1.5 hours cover the basics. Prerequisites include:
1. Basic skills: Flowchart/swimlane diagramming; experience with event-triggered tools (Axure, IFTTT, iOS Shortcuts, RPA). 2. Data basics: Data types (numbers, dates, strings, arrays, binary files) and JSON. 3. API concepts: GET/POST, API keys, cURL — no deep technical knowledge needed. 4. Prompting knowledge: system prompt vs user prompt for AI chat nodes. 5. Agent structure: Chat + Memory + Tool for configuring Agent nodes.
Use cases: automated social media posting and customer email replies, batch image generation/editing, and cloud file management (invoice organization, document classification). Note that better tools may exist for some scenarios.
Ecosystem contrast: Overseas markets enjoy a thriving open-API ecosystem with companies and intermediaries of all sizes; domestic (China) users tend to rely on full-suite services from large platforms.