Author HubMatthias Bastian

Matthias is the co-founder and publisher of THE DECODER, exploring how AI is fundamentally changing the relationship between humans and computers.
AI research

Wait a minute! Researchers say AI's "chains of thought" are not signs of human-like reasoning

AI in practice
Update

US Copyright Office says fair use does not cover AI trained on "vast troves of copyrighted works

AI research

OpenAI's o3 is less AGI than originally measured

Ad
AI and society
Short

U.S. President Donald Trump has dismissed Shira Perlmutter, head of the U.S. Copyright Office, shortly after her office released a report opposing broad "fair use" exemptions for AI training purposes. The report's stance conflicts with the interests of Trump ally Elon Musk and much of the AI industry. Perlmutter had served since 2020. Rep. Joseph Morelle, the top Democrat on the House Committee on Administration, called the dismissal an "unprecedented power grab with no legal basis" and said the timing was "certainly no coincidence."

This action once again tramples on Congress’s Article One authority and throws a trillion-dollar industry into chaos.

Rep. Joe Morelle

Ad
Ad
AI in practice
Short

Reasoning tasks sharply raise AI costs, according to a new analysis by Artificial Analysis. Google's Gemini Flash 2.5 costs 150 times more to run than Flash 2.0, due to using 17 times more tokens and charging $3.50 per million output tokens with reasoning, compared to $0.40 for the earlier model. This makes Flash 2.5 the most expensive model in terms of token use for logic. OpenAI's o4-mini costs more per token but used fewer tokens overall, making it cheaper in the benchmark.

Bar chart titled “Cost to Run Artificial Analysis Intelligence Index.” It shows total U.S. dollar costs to complete all tests in the Artificial Analysis Intelligence Index using different AI models. Bars are split into three colors: Input (blue), Reasoning (purple), Output (green). On the left are the most expensive models: GPT-3 ($1,951), Claude 3 Opus ($1,485), Gemini 2.5 Pro ($844). In the middle: Gemini 2.5 Flash with reasoning ($445), o4-mini (high) ($323). On the right are the cheapest models: Gemini 2.0 Flash ($3), Llama 3 8B ($2). A purple arrow above highlights the cost gap between Gemini 2.0 Flash and Gemini 2.5 Flash with reasoning, labeled “150x.” Source: Artificial Analysis.
Google's Gemini Flash 2.5 costs 150 times more to run with reasoning enabled than Flash 2.0, due to higher token use and pricing. | Image: Artificial Analysis
Ad
Ad
AI in practice
Short

Google introduces "implicit caching" in Gemini 2.5, aiming to cut developer costs by as much as 75 percent. The new feature automatically detects and stores recurring content, so repeated prompts are only processed once. According to Google, this can lead to significant savings compared to the old explicit caching method, where users had to set up their own cache. To maximize the benefits, Google recommends putting the stable part of a prompt—like system instructions—at the start, and adding user-specific input, such as questions, afterwards. Implicit caching kicks in for Gemini 2.5 Flash starting at 1,024 tokens, and for Pro versions from 2,048 tokens onwards. More details and best practices are available in the Gemini API documentation.

AI in practice
Short

OpenAI is rolling out a new GitHub connector for ChatGPT's deep research agent. Users with Plus, Pro, or Team subscriptions can now connect their own GitHub repositories and ask questions about their code. ChatGPT searches through the source code and documentation in the repo, then returns a detailed report with source references. Only content that users already have access to is visible to ChatGPT, so existing permissions apply. The connector will become available to users over the next few days, with support for enterprise customers coming soon. According to OpenAI Product Manager Nate Gonzalez, the goal is to better integrate ChatGPT into internal workflows. OpenAI also plans to add more deep research connectors in the future.

Ad
Ad
AI in practice
Short

The Trump administration is moving to overturn the Biden administration's AI chip export restrictions, according to a spokesperson for the US Department of Commerce. Biden's rules, which were set to take effect on May 15, would have split the world into three different zones for AI chip exports, capping shipments to most countries. Trump officials say the current system is "overly complex, overly bureaucratic," and want to replace it with "a much simpler rule that unleashes American innovation and ensures American AI dominance." One idea on the table is a global licensing system based on intergovernmental agreements, but there is no set timeline for new regulations yet. Since taking office, the Trump administration has already rolled back several other Biden-era AI policies.

Google News

Bank details

IBAN: DE87 1203 0000 1086 0070 75
Account holder: DEEP CONTENT GbR
Purpose: Support THE DECODER
AI research

New Othello experiment supports the world model hypothesis for large language models

AI in practice

ChatGPT might be draining your brain, MIT warns - what ‘cognitive debt’ means for you

AI research

Meta's latest model highlights the challenge AI faces in long-term planning and causal reasoning

Google News