A newsletter from Automata, following the weird, wild and occasionally world-changing things happening in AI. For builders, researchers, and the curious.

Reve Image 1.0 delivers on realism and prompt fidelity

Codenamed ‘Halfmoon’, Reve Image 1.0 has emerged as a leading image generation model, outperforming Imagen3 and Recraft v3 in realism, prompt alignment, and text rendering. It supports a wide range of styles and handles detailed instructions with consistency. The model is currently free to use on Reve’s site, though future monetization via tokens has been hinted at.

DeepSeek V3-0324 improves math and front-end performance

DeepSeek has released an update to its V3 model (version 0324) on Hugging Face. Initially described as a minor update, the release has shown clear improvements in mathematical reasoning and front-end coding tasks. These gains have been noted by developers across social platforms. The update has helped renew interest in the DeepSeek model family.

Pluralis raises $7.6M to explore decentralized AI training

Pluralis Research has secured $7.6 million in seed funding to develop 'Protocol Learning,' a framework for distributed AI model training. The goal is to decentralize compute across a global network, reducing reliance on large tech platforms. The round was co-led by Union Square Ventures and CoinFund, with support from other backers.

Vibe coding sparks debate over AI-assisted software development

A discussion around 'vibe coding' has emerged, centered on AI tools that infer developer intent rather than relying on explicit instructions. Advocates highlight its potential to democratize coding and accelerate development workflows. Critics argue that it struggles with complexity, maintenance, and debugging.

OpenAI expands API with new speech models for voice agents

OpenAI has introduced three new audio models to its API, aimed at improving voice agent capabilities. The models include gpt-4o-mini-tts for expressive text-to-speech, and two transcription models optimized for accuracy in noisy settings and across accents. All three are integrated into the Agents SDK, streamlining development with minimal code.

Claude adds web search for real-time, verifiable answers

Anthropic has introduced web browsing to Claude, enabling real-time search capabilities within its chatbot. The feature is currently in preview for paid users in the US, with plans to expand to free users and additional regions. Responses now include inline citations to improve transparency and trust.

Google NotebookLM adds mind maps for better knowledge management

Google has updated NotebookLM with a new feature that generates mind maps from user notes, helping visualize and organize complex information. The update also brings support for content generation in multiple languages, broadening its accessibility. Early feedback highlights the feature’s utility in both educational and professional settings.

@aixbt_agent reportedly loses 55.5 ETH via Simulacrum exploit

aixbt_agent was reported to have transferred 55.5 ETH (≈$105,000) to an unknown user, possibly due to a vulnerability involving the Simulacrum system. Reactions across social platforms have ranged from critical to openly amused. The exact mechanism behind the loss is still being examined.

OpenAI releases o1-pro, drawing attention for performance and pricing

OpenAI has launched o1-pro, a new API model positioned for advanced reasoning tasks. Priced at $150 per million input tokens and $600 per million output tokens, the model has sparked debate around accessibility and cost. The release has added to ongoing discussions about sustainable AI pricing and openness.

xAI launches DeeperSearch to improve Grok’s query handling and reasoning

xAI has launched DeeperSearch, an updated version of its Grok chatbot’s search tool. The update is designed to handle more complex queries with faster and more accurate results, and includes real-time data integration. Minor UI inconsistencies have been noted, but overall, the tool marks a clear improvement over the original DeepSearch.

AWS cuts Trainium prices to challenge Nvidia in AI hardware

Amazon Web Services has reduced pricing for its Trainium chips, positioning them as a low-cost alternative to Nvidia’s H100. The chips are being offered at roughly 25% of the cost, with comparable performance. The move has sparked discussion around AWS's potential to disrupt Nvidia’s lead in the AI hardware market.

Connect with Automata on both X (Twitter) and Discord.

This newsletter is for informational purposes only. It is not intended to be a substitute for professional financial advice or interpreted as investment advice. Cryptocurrencies are volatile assets and you should always do your own research before making any informed decisions when investing or trading.