AI Advances: TabPFN-3 Outshines Traditional ML and Claude Opus 4.8 Enhances Autonomous Work

Prior Labs' TabPFN-3 and Anthropic's Claude Opus 4.8 are making significant strides in the AI landscape, outperforming traditional machine learning models and enhancing autonomous work capabilities, respectively.

TabPFN-3 Beats Tuned Classical ML

Prior Labs introduces TabPFN-3, a new model that surpasses tuned classical machine learning algorithms with a 90% win rate. This breakthrough simplifies the process of handling structured data, eliminating the need for extensive tuning and configuration.

Claude Opus 4.8 Improves Autonomous Work

Anthropic releases Claude Opus 4.8, which includes a 2.5x faster mode and parallel subagents. The new version is more adept at recognizing its limitations, reducing unsupported claims and improving the reliability of autonomous tasks. It supports a 1M token context window and 128k max output tokens, available on the API, Claude Code, and major cloud platforms.

NVIDIA Invests $150B in Taiwan

NVIDIA announces a $150 billion investment in Taiwan to solidify the region as the epicenter of the AI revolution. The move aims to expand partnerships with TSMC and other local tech companies, leveraging advanced packaging technology not yet available in the U.S. This investment underscores NVIDIA’s commitment to expanding the AI ecosystem and boosting its bottom line.

Anthropic and OpenAI Find Product-Market Fit

Anthropic and OpenAI have found their product-market fit with coding and general-purpose agent products. Both companies are now aggressively pricing their APIs, with customers spending over $200 per month per user, significantly covering their costs compared to lower pricing tiers. Coding agents are driving this increased spending.

MLXcel Goes Open Source

MLXcel, an inference engine built for Apple Silicon, is now open source under the Apache 2.0 license. On M5 Max, MLXcel reaches up to 2.70x the prefill median of mlx-lm and matches it on decode. The supported-models list spans over 70 text architectures and 22 vision-language models (VLMs). This move aims to democratize AI inference by making it accessible on a broader range of devices.

Study Reveals Secrets to Better MoE Models

A study of 2,000 training runs provides insights into building better mixture-of-experts (MoE) models. The research helps in designing efficient architectures before committing computational resources, ensuring that most of the model remains inactive until needed, thus optimizing performance and cost.

References

• AI News Briefs BULLETIN BOARD for May 2026 | Radical Data Science