Author HubMaximilian Schreiner

Max is the managing editor of THE DECODER, bringing his background in philosophy to explore questions of consciousness and whether machines truly think or just pretend to.
AI and society

OpenAI restructures under new foundation, Microsoft takes 27 percent stake

AI research

OpenAI outperforms humans and Google at the world's top collegiate programming contest

AI in practice

New data from OpenAI and Anthropic show how people actually use ChatGPT and Claude

Ad
AI in practice
Short

Google's MatCha is a foundation model trained for both chart de-rendering and mathematical reasoning. Chart de-rendering explores the reverse engineering of charts, plots, or graphics to reveal their underlying data table or code, while math reasoning seeks to solve question-based problems on textual mathematical datasets. By combining these tasks, MatCha significantly outperforms existing models for visual language understanding of charts. The researchers also proposed DePlot, a model built on top of MatCha for improved reasoning on charts through translation to tables.

Bild: ChartQA
Ad
Ad
AI in practice
Short

Is GPT-4 getting worse? Peter Welinder, VP Product at OpenAI, comments on the rumors: "No, we haven't made GPT-4 dumber. Quite the opposite: we make each new version smarter than the previous one," says Welinder. His hypothesis: "When you use it more heavily, you start noticing issues you didn't see before."

Still, he asked on Twitter for examples of where users felt GPT-4's performance had regressed. OpenAI will look into this, Welinder promises. One example from the comments, where GPT-4 only gives the correct answer after a second attempt, he claims is a bug.

Field reports of GPT-4 performance degradation have been around for a few weeks now, often pointing to a performance difference between access via API and via ChatGPT, where the OpenAI language model integration is regularly adjusted. The opacity of this process may also contribute to uncertainty among users.

Ad
Ad
AI in practice
Short

AI tool RFdiffusion, a neural network developed by the University of Washington, can rapidly design custom proteins with potential applications in vaccines, therapeutics, and biomaterials. The tool borrows principles from neural networks that generate realistic images, such as Stable Diffusion, DALL-E, or Midjourney, creating complex and diverse protein shapes when it "denoises" a random assortment of amino acids.

Although the designs show promise in early experimental tests, researchers are still working to improve the AI's ability to design proteins with more complex active sites and specific reactions. RFdiffusion is available on GitHub.

AI research
Short

A team of Chinese researchers has developed an AI tool capable of designing a working RISC-V CPU in just five hours. It successfully ran the Linux operating system and performed comparably to the human-designed Intel 80486SX CPU from 1991. According to the team, with a working chip in just under five hours, the AI was able to design the chip 1,000 times faster than a human team. The AI was also able to autonomously make discoveries in computer design, uncovering the von Neumann architecture first invented in 1945.

The authors suggest that their approach, outlined in a study titled "Pushing the Limits of Machine Design: Automated CPU Design with AI" could revolutionize the semiconductor industry by significantly shortening the design cycle.

Ad
Ad
Google News

Bank details

IBAN: DE88 2507 0070 0053 0014 00
BIC: DEUTDE2HXXX
Account holder: Deep Content GmbH
Purpose: Support THE DECODER
AI research

The ARC benchmark's fall marks another casualty of relentless AI optimization

AI in practice

DeepseekMath-V2 is Deepseek's latest attempt to pop the US AI bubble

AI research

Frustrated authors withdraw papers after realizing their reviewers are just lazy LLMs

Google News