☕️ The Concerns & Consequences of GPT-4o's 'Too Good' Image Generation

Just 2 weeks after the release of Google's next-gen image generator, OpenAI debuted its own as part of a major upgrade to bring native image generation to ChatGPT. Other key highlights include:

• Google releases Gemini 2.5, a next-gen upgrade to its Gemini series of reasoning models

• DeepSeek's V3 becomes one of the most competitive non-reasoning models following a recent update

• Microsoft brings deep research tools to its Copilot chatbot with the addition of two new modes

Join us at AI Tangle as we untangle this week's happenings in AI!

OpenAI brings native image generation capabilities to its GPT-4o model

During a livestream on Tuesday, OpenAI CEO Sam Altman announced a major upgrade for ChatGPT, enabling it to create and modify images using GPT-4o at a higher quality than ever before. GPT-4o's image generation provides a new standard for GenAI image quality and, most importantly, highly accurate text rendering. In a statement to the Wall Street Journal, OpenAI CFO Brad Lightcap said that respecting artists' rights and avoiding style mimicry of living artists was an important focal point. The features are available to API developers, Pro users, Plus users, and eventually Free users, the latter of which has been delayed due to unexpectedly high demand for perhaps the wrong reasons.

​

​What is happening with GPT-4o, the internet, and Studio Ghibli?​

Not long after news of GPT-4o's image generation capabilities took off, it sparked a trend on social media where users began to emulate the style of Japanese animation company Studio Ghibli, one of the most appraised and esteemed in recent history. Though OpenAI has seemingly reinforced its content policies as of writing, it nonetheless sparked an equally fierce debate about AI's ever-controversial role in the field, especially when considering the stance of Studio Ghibli co-founder Hayao Miyazaki, who fervently voiced his disgust for the technology in the creative sector, going as far as to call it an "insult to life itself" back in 2016.

Google pushes the boundaries once more with Gemini 2.5, the company's latest flagship reasoning model

Google has continued to step up the pace in the AI game, now announcing the release of Gemini 2.5, an updated generation of its reasoning AI models to tackle more complex queries. Gemini 2.5 readily outperforms or equals competitors on common benchmarks for understanding, mathematics, and, most importantly, coding, where it has dethroned Anthropic's previously incredibly dominant Claude 3.7 Sonnet. It comes with native multimodality and an enormous 2 million token context window, which many believe the company has its investments in TPU technology to thank for helping it make possible at such a low price.

​

DeepSeek's latest update to its non-reasoning V3 model propels it to the top of the competition

Earlier this week, DeepSeek unveiled an upgraded version of its DeepSeek V3 LLM via Hugging Face, though the upgrade ended up being more impactful than many may have expected. The new model shows marked improvements in reasoning and coding, bolstering DeepSeek's competitive edge against Western rivals like OpenAI and Anthropic. Most importantly, however, DeepSeek V3 now claims the spot as the "best" non-reasoning model on top of being open-source and significantly cheaper than OpenAI's GPT-4.5.

​

Microsoft infuses Copilot with two new modes for deep research-like functionality

Following the likes of Perplexity, OpenAI, and Google, Microsoft is next in line to add "deep research"-like tools to its Copilot AI chatbot. In Microsoft's case, the company is giving Copilot two separate modes, Researcher and Analyst. Researcher integrates OpenAI's deep research model with "advanced orchestration" and "deep search capabilities," while Analyst, built on OpenAI's o3-mini model, is optimized for advanced data analysis and can also run Python code for more difficult queries. The features will be available starting in April for participants of its new Frontier program.

​

Anthropic secures an early win as music publishers fail to prove Anthropic's conduct causing "irreparable harm"

Popular AI startup Anthropic has secured itself an early victory as it convinced a California federal judge to reject a broad preliminary bid by music publishers. The bid, presented by publishers such as Universal Music Group, was presented to block the company from using copyrighted lyrics to train its Claude AI chatbot. However, US District Judge Eumi Lee found that the music publishers failed to demonstrate "irreparable harm" from Anthropic's conduct, though they remain confident in their broader case.

​

CoreWeave slashes $4 billion IPO expectations down to $1.5 billion amid waning Wall Street interest

New Jersey-based cloud computing provider CoreWeave has significantly slashed its expectations after reducing the size and value of its planned IPO as investor demand for AI infrastructure wanes on Wall Street. The company is saying it now only expects to raise about $1.5 billion by selling around 37.5 million shares at $40 each, a stark contrast to its original $4 billion goal. Nvidia, an already major stakeholder, is said to be at the crux of the IPO by anchoring it at that $40 price point and is set to invest an additional $250 million in the company.

​

Ex-Meta AI executives depart from the tech giant to launch an AI agent startup with $15 million in fundraiser money

Two former Meta AI executives have reportedly raised $15 million for Yutori, a San Francisco-based startup developing autonomous AI personal assistants. Led by Rob Toews at Radical Ventures with participation from investors including Felicis, "AI godmother" Fei-Fei Li, and Google DeepMind chief scientist Jeff Dean, Yutori says it wants to redefine how users interact with AI agents. Right now, the company is focusing on post-training models to improve web navigation and performance, taking aim at tasks ranging from food orders to complex travel logistics.

​Substrata - Substrata in an AI-based deal-closing tool that analyzes analyzes real-time human dynamics and intentions to help dealmakers sell smarter and close more.

​Thunderbit - Thunderbit is a versatile, no-code, out-of-the-box AI toolkit with a web assistant that empowers business users to readily build and deploy AI tools and automation products in seconds.

​Supernormal - Spend less time writing, polishing, and sharing notes and more time on the work that matters with the help of Supernormal's AI-powered meeting transcriptions.

Emboldened by Trump, AI Companies Lobby for Fewer Rules (8-min read)

From Meta to OpenAI, tech giants have swung the other way from Democrat-led times and are now urging the US to ease AI regulations instead, with many arguing that fewer restrictions are needed to help the new and on-the-rise transformative industry get up and running.