Welcome back to Daily Zaps, your regularly-scheduled dose of AI news ⚡️ 

Here’s what we got for ya today:

• 🤖 OpenAI releases “deep research” for ChatGPT

• ❌ EU bans AI systems deemed ‘unacceptable risk’

• 💰 DeepSeek’s total hardware spend is well over $500 million

• ☔️ Google introduces AI agent that aces 15-day weather forecasts

Let’s get right into it!

Sign Up | Advertise | Tools

STARTUPS

OpenAI releases “deep research” for ChatGPT

OpenAI has unveiled ChatGPT deep research, a new AI agent designed to help professionals and consumers conduct detailed, multi-source research rather than relying on quick summaries. Targeted at users in fields like finance, science, policy, and engineering—and even those making significant purchases—the tool is initially available to ChatGPT Pro users with a monthly query limit, with future expansions planned for other tiers.

By leveraging an enhanced version of its o3 reasoning model, which has been trained through reinforcement learning on real-world tasks, deep research provides fully documented, text-only outputs complete with citations, although it still faces challenges such as occasional inaccuracies, formatting errors, and difficulty in distinguishing authoritative information. OpenAI intends to further improve the feature by adding visualizations and support for specialized data sources, positioning it as a more robust alternative to simple chatbot summaries amid rising competition, including similar offerings from Google.

GLOBAL NEWS

EU bans AI systems deemed ‘unacceptable risk’

As of February 2, the European Union has begun enforcing the first compliance deadline under its AI Act, allowing regulators to ban AI systems deemed to pose an “unacceptable risk.” The Act categorizes AI into four risk levels, with minimal and limited risk systems facing little to no regulation, high-risk AI subject to strict oversight, and unacceptable-risk AI prohibited entirely. Banned applications include AI for social scoring, subliminal manipulation, crime prediction based on appearance, real-time biometric surveillance, and emotion inference in workplaces and schools. Companies found violating these restrictions could face fines of up to €35 million or 7% of their annual revenue.

While over 100 companies, including Amazon and Google, voluntarily pledged compliance through the EU AI Pact last year, major players like Meta, Apple, and Mistral did not sign. Some exceptions exist, such as biometric data collection for law enforcement in specific cases with authorization, or emotion inference for medical and safety purposes. The European Commission is expected to release further guidelines in early 2025, as organizations navigate compliance complexities and interactions with other regulations like GDPR and NIS2.

STARTUPS

DeepSeek’s total hardware spend is well over $500 million

China’s AI startup DeepSeek made headlines this week with its R1 model, drawing industry attention over its reported $6 million training cost—a figure that reflects only the official training phase, excluding prior research, experiments, and infrastructure investments. Analysts from SemiAnalysis estimate that DeepSeek’s total hardware spend is well over $500 million, considering research, synthetic data generation, and operational costs.

Despite U.S. chip export restrictions, DeepSeek R1 has been praised for its reasoning capabilities, quickly rivaling OpenAI’s o1 model, and even surpassing ChatGPT as the most-downloaded free app on Apple’s U.S. App Store. The open-source nature of R1 has fueled concerns over AI competition, with OpenAI CEO Sam Altman acknowledging its strength while alleging that DeepSeek may have improperly used OpenAI’s data.

BIG TECH

Google introduces AI agent that aces 15-day weather forecasts

Google DeepMind has introduced GenCast, an AI model that outperforms traditional weather forecasting systems, predicting weather conditions with 97% accuracy compared to the Ensemble Forecast (ENS). Unlike previous AI-driven models, GenCast relies solely on deep learning, predicting weather similarly to how ChatGPT predicts text. Trained on 40 years of weather data, it excels in forecasting wind patterns and extreme weather events, aiding in disaster preparedness and renewable energy management.

Unlike deterministic models like Huawei’s Pangu-Weather, which provide fixed predictions, GenCast generates probabilistic forecasts, offering likelihoods for various weather outcomes. However, it still depends on historical datasets like ERA5, and its accuracy diminishes over time without updated training. While AI models struggle with certain atmospheric conditions, meteorologists remain crucial for interpreting forecasts and integrating additional data. DeepMind plans to refine GenCast by incorporating real-time observational data, aiming for more precise and reliable weather predictions.