Sakana AI and 360 Group launch frontier models into the gap left by export controls
Asian AI startups Sakana AI and 360 Group launched frontier-class models to fill the gap left by U.S. export controls on Anthropic’s Mythos 5, while OpenAI retired GPT-4.5 from ChatGPT as the GPT-4 era formally ends. Modular shipped Apple silicon GPU support for its MAX inference engine, and Sebastian Raschka’s benchmarks showed open-weight local coding agents achieving GPT-5.5-level performance on consumer hardware.
Model Releases #
Asian AI startups launch Mythos-like models as Anthropic’s export ban drags on #
TechCrunch / Hacker News (236 points)
Tokyo-based Sakana AI unveiled Fugu, claiming frontier-class capabilities “shoulder-to-shoulder” with restricted U.S. models, while Chinese cybersecurity firm 360 Group released Tulongfeng, targeting Mythos-level vulnerability detection. Both companies framed their launches explicitly as alternatives to fill the gap created by the Trump administration’s export controls on Anthropic’s Mythos 5 and Fable 5, positioning non-U.S. AI capabilities as strategic assets free from export risk.
Developer Tools #
OpenAI retires GPT-4.5 from ChatGPT #
OpenAI / TechRadar
OpenAI retired GPT-4.5 from ChatGPT on June 27, completing the sunset of the entire GPT-4 generation that defined the 2023-2025 era of commercial AI. Existing conversations automatically migrate to GPT-5.5, and the change applies only to ChatGPT – the API is unaffected. The retirement marks the end of a model family that catalyzed the current AI industry, with GPT-5.5 now the baseline consumer offering.
Using Local Coding Agents #
Ahead of AI (Sebastian Raschka)
Sebastian Raschka evaluated open-weight models like Qwen3.6 35B-A3B and North Mini Code 1.0 running in local coding harnesses (Qwen-Code, Codex, Claude Code) as alternatives to proprietary cloud subscriptions, finding performance comparable to GPT-5.5 at 30-40 tokens/second on consumer hardware. Codex used the fewest tokens while matching Claude Code’s task success rates, making it the most efficient harness choice in the evaluation. For teams managing AI coding budgets, the findings suggest that local open-weight setups now offer a viable alternative to cloud-based coding agents with fixed costs and full privacy.
MAX models can now run on Apple silicon GPUs #
Modular / Lobsters AI
Modular’s MAX 26.4 and current nightly builds now support running text LLMs, vision models, and image diffusion models on M1-M5 Apple silicon GPUs, expanding beyond CPU-only inference for the first time. Users can run models like Qwen3.5-0.8B or FLUX.2-klein directly on Macs with sufficient RAM via simple CLI commands. The release broadens the local inference ecosystem on Apple hardware alongside MLX, giving developers another option for on-device model serving.
Research & Papers #
Comparing Transformers and Hybrid Models at the Token Level #
Yanhong Li, William Merrill / Lobsters AI
This paper demonstrates that while transformers and hybrid architectures (combining attention with recurrent mechanisms) achieve similar aggregate performance, they process information fundamentally differently at individual token positions. Token-level analysis reveals distinct processing dynamics that aggregate benchmarks cannot capture, suggesting that architectural differences manifest in how models attend to and process sequential information. For practitioners choosing between architectures, the finding means that benchmark parity does not imply behavioral equivalence – the choice of architecture still matters for specific use cases even when headline scores converge.
AI Learns the “Dark Art” of RF Chip Design #
IEEE Spectrum / Lobsters AI
Researchers at Princeton led by Kaushik Sengupta applied reinforcement learning and diffusion models to radio-frequency integrated circuit design, accelerating a process that traditionally takes years of manual expert iteration down to days or minutes. The AI approach uses convolutional neural networks to emulate electromagnetic behavior without traditional simulation, achieving superior performance metrics in power amplifiers for 5G, autonomous vehicles, and satellite communications. The practical barrier remains data access – proprietary RF design datasets are locked behind NDAs, limiting the development of universal models.
Security #
Post-Mythos Cybersecurity: Keep calm and carry on #
Cephalosec / Hacker News (155 points)
This widely-discussed article examines the cybersecurity implications of Mythos-class AI models, arguing for a measured defensive response over panic. The piece generated significant discussion among security practitioners on Hacker News about adapting defensive strategies to an era where frontier AI models can autonomously identify and exploit vulnerabilities. The full article was unavailable for detailed analysis at fetch time (HTTP 403).
Funding & Business #
Apple Vision Pro exec is reportedly leaving for OpenAI #
TechCrunch
Paul Meade, Apple’s vice president overseeing Vision Pro and smart glasses development, is departing to join OpenAI’s hardware division, where the company is collaborating with former Apple design chief Jony Ive on a device described as “more peaceful and calm than an iPhone.” The hire signals that OpenAI’s hardware ambitions are serious enough to attract top-tier talent from Apple’s most advanced hardware programs. The departure follows John Ternus’s elevation to Apple CEO, suggesting internal restructuring may be driving leadership changes in Apple’s hardware organization.
Infrastructure #
AMD Strix Halo RDMA Cluster Setup Guide #
GitHub / Hacker News (145 points)
A community guide for building RDMA-connected clusters using AMD Strix Halo consumer processors to run vLLM inference received significant attention on Hacker News. The guide documents how to set up high-speed interconnects between consumer AMD systems for distributed LLM serving, an approach that bypasses the cost of enterprise GPU clusters. For teams exploring low-cost inference infrastructure, consumer AMD hardware with RDMA networking represents a viable alternative for vLLM deployments at a fraction of data center GPU costs.
SoftBank’s CEO isn’t the only one with questions about Elon Musk’s orbital data center hype #
TechCrunch
SoftBank CEO Masayoshi Son and OpenAI’s Sam Altman have publicly dismissed Elon Musk’s orbital data center vision, arguing it “won’t do much to cut costs and will take too long” when the AI industry needs compute solutions within years, not decades. Critics note that building satellite constellations requiring frequent replacement primarily generates launch business for SpaceX itself, raising questions about whether the proposal serves SpaceX’s commercial interests more than the industry’s actual compute shortage. The skepticism from major AI investors and operators suggests the orbital data center concept faces fundamental economic challenges beyond the engineering ones.
Open Source #
GPT2-BASIC: Portable Machine Intelligence in BASIC #
GitHub / Lobsters AI
A community project implementing GPT-2 inference entirely in BASIC, demonstrating that transformer inference can run on minimal, decades-old computing environments without modern ML frameworks. While not production-relevant, the project serves as an educational tool illustrating the mathematical simplicity underlying transformer architectures when stripped of optimization layers. For developers learning ML fundamentals, it provides a uniquely transparent view of how transformer inference works at the arithmetic level.
Threads to Watch #
Export controls are creating a parallel AI ecosystem in Asia. Sakana AI and 360 Group launched frontier-class models explicitly targeting the gap created by U.S. restrictions on Mythos 5, framing non-American AI as strategically preferable. Combined with yesterday’s news that the Mythos ban was partially lifted for 100+ U.S. entities, the pattern suggests that export controls are simultaneously constraining global access to U.S. models and accelerating the development of non-U.S. alternatives – potentially fragmenting the frontier model landscape along geopolitical lines.
Local AI inference is reaching practical parity with cloud offerings. Sebastian Raschka’s evaluation shows open-weight models matching GPT-5.5 at 30-40 tokens/second on consumer hardware, Modular shipped Apple silicon GPU support for MAX, and the AMD Strix Halo RDMA cluster guide demonstrates consumer-grade distributed inference for vLLM. Together, these developments suggest that the cost and privacy advantages of local inference are becoming accessible to individual developers, not just organizations with dedicated infrastructure teams.