OpenAI continues to make the headlines with a massive week of releases, from a prototype social media platform to a coding interface to the debut of its long-awaited o3 and o4-mini reasoning models.
Other key takeaways include:
- Google releases a preview version of its ultra-cheap Gemini 2.5 Flash AI model for developers
- OpenAI could have an AI-centric social media platform in the works to compete against bitter rival X
- Nvidia braces for a $5.5 billion hit to its revenue as export controls tighten around its H20 chip
Join us at AI Tangle as we untangle this week's happenings in AI!
|
As part of its major release week, OpenAI finally launched the standalone commercial version of o3, its most advanced model yet, and o4-mini reasoning models on Wednesday. OpenAI says that the pair are the company's first models that can "interpret" images through its elaborate chain-of-thought process and, unlike its previous reasoning models, can generate responses using ChatGPT's web browsing, execution of Python code, and image processing - and they're already available for subscribers to OpenAI's Pro, Plus, and Team plans.
How do o3 and o4-mini hold up against the industry standard?
Along with an additional variant of o4-mini called o4-mini-high that focuses on increased reliability, the new o3 and o4-mini models fiercely compete with Google's industry-defining Gemini 2.5 Pro, trading blows across a variety of benchmarks like SWE, MMMU, AIME 2024/2025, Humanity's Last Exam, GPQA, and more, but ultimately coming out either on par or on top by a small margin.
However, despite the results, the main problem of o3 and o4-mini lies in their affordability in comparison to Gemini 2.5 Pro, as exposed by the Aider Polyglot benchmark. Developers still willing to put up with the costs, however, will be happy to hear that OpenAI launched a new, open source command-line coding agent called Codex CLI in parallel with the launch of o3 and o4-mini with both models already available on it.
|
Join Mark Hinkle for a live, hands-on workshop session tailored for business professionals aiming to enhance their AI interactions. Discover how to craft effective prompts that yield tangible business results, from generating marketing copy to using AI to generate prompts that get results.
From this workshop, you'll learn:
- Effective prompt structures for diverse business applications
- Prompt optimization via live demonstrations
- Techniques to develop scalable and repeatable prompts
The workshop takes place on the 29th of April at 6:30 PM ET - don't miss out on a free lesson to use AI better and register now!
|
Just a day after OpenAI debuted its o3 and o4-mini continuing to struggle against the affordability of Gemini 2.5 Pro, Google lowered its price point even further with the release of Gemini 2.5 Flash. Available as a preview for developers via Google AI Studio and Vertex AI, Gemini 2.5 Flash builds on its low-latency 2.0 model. As Google's "first fully hybrid reasoning model," the model comes with a built-in toggle that allows developers to turn Gemini's "thinking" on or off, depending on the needs of the query, and set token budgets to balance quality, cost, and latency.
OpenAI is reportedly developing a platform akin to X/Twitter, with the company working on a social feed with an explicit focus on ChatGPT's image generation. The project is still in its prototype stage as Sam Altman allegedly circulates information for outsider feedback. As of now, it is unclear whether the project will launch as a separate app or eventually become part of ChatGPT. Regardless of what route it takes, the move will very likely increase tensions between OpenAI and Elon Musk further, along with putting it in closer competition with Meta.
Nvidia has warned that it expects to take a $5.5 billion hit to its revenue this quarter due to new US export controls on its designed-for-China H20 chip. Despite the administration reversing course on plans to restrict the chip just last week, Nvidia's H20 now requires a special license for sale, as US authorities continue to prevent any US-made advanced AI chips from ending up in Chinese supercomputers. Soon after the news on late Tuesday, Nvidia shares were sent tumbling down roughly 6% in after-hours trading, also affecting other chipmakers on a global scale.
As detailed in a blog post, Anthropic has announced a deeply-wired integration with Google's productivity suite, bringing its Claude chatbot to Google Workspace by connecting with common Google apps like Gmail, Google Calendar, and Google Docs. Additionally, the company also launched Claude Research, a feature similar to the likes of Google's, OpenAI's, or Perplexity's own takes on deep web research for generating highly detailed responses. Both features are rolling out as part of a beta to Anthropic's Max, Team, Enterprise, and Pro plan subscribers.
The Wikimedia Foundation has partnered with Google‑owned data science platform Kaggle to release a dataset of structured Wikipedia content to dissuade AI bots from scraping the platform directly for training data. The dataset, available in English and French, provides openly licensed research summaries, short descriptions, image links, infobox data, and article sections while omitting citations and non‑written elements like audio. By offering a clean alternative to raw article scraping, Wikimedia hopes the publicly accessible dataset will fend off bot scrapers and ease the strain on its aching servers.
OpenAI has entered preliminary talks to acquire Windsurf, an AI-assisted coding tool formerly known as Codeium, for around $3 billion, according to Bloomberg, marking what could potentially be OpenAI's largest deal yet. With OpenAI having roughly $40 billion in funding to splurge, the move is in line with last year's acquisition of Rockset as it looks to expand further into developer tools and enterprise infrastructure. Before Windsurf was considered, however, OpenAI may have had its eyes on acquiring fellow competitor Anysphere's Cursor first.
|
Kraftul - Kraftul is the copilot for product teams that help businesses' get insights into what makes their products work and analyzes support tickets, calls, and reviews to more easily understand and condense user feedback.
Hoop - Made for the busy professional, Hoop is an AI task management platform that captures and prioritizes your tasks with a global task list across your teams.
Trupeer - Trupeer turns your simple screen recordings into polished videos and detailed guides, generating transcripts and voiceovers with support for multiple formats and customization options.
|
| What did you think of this newsletter? Let us know! |
|
|
|
|
|