After months of work, delays, and 200,000 GPUs at xAI's Memphis data center, Grok 3 finally sees the light of day, leaving just one question - how does it hold up? Other key highlights include:
- Musk-led group's nearly $100 billion bid to take over OpenAI gets rejected by its board of directors
- AI search company Perplexity joins the deep research tool race with a release of its own
- South Korea puts new DeepSeek downloads on pause after a sudden uproar in popularity
Join us at AI Tangle as we untangle this week's happenings in AI!
|
After a delay following an optimistic release deadline estimate of late 2024, the long overdue Grok 3 model from Elon Musk's xAI team finally saw the light of day this Monday. The model was unveiled during a livestream presentation featuring members of the xAI team and Elon Musk himself. After months of work, delays, and 200,000 GPUs at xAI's massive Memphis data center powering its training, what are the results?
How does xAI's Grok 3 fare against the world?
First available to X's Premium+ subscribers with additional features gated behind a new subscription called SuperGrok, Grok 3 comes in two different sizes: the mainstay Grok 3 and the smaller Grok 3 mini. Both of the models come with optional modes for reasoning called "Think" and "Big Brain," depending on how difficult users believe their query is. However, users may notice that some of the reasoning models' "thoughts" may be obscured, which Musk stated is to ward off distillation practices, something that DeepSeek is allegedly guilty of.
The crowdsourced AI model ranking platform Chatbot Arena has an early version of Grok 3 named "chocolate" as the all-around #1 on the platform - important to note is that o3 is currently not present on the arena. xAI itself reports that both Grok 3 and Grok 3 mini either match or beat out their respective non-reasoning and reasoning counterparts, which includes the likes of Gemini 2.0 Pro, o3-mini-high, o1, DeepSeek-R1, Claude 3.5 Sonnet, and more, on the AIME 2024/25 and GQPA benchmarks, even though there is no current way to verify these results.
|
After a Musk-led investor group offered nearly $100 billion to take over the startup, perhaps unsurprisingly, OpenAI's board of directors unanimously shut down the proposal. Chairman Bret Taylor said in a statement on Friday last week that "OpenAI is not for sale, and the board has unanimously rejected Mr. Musk's latest attempt to disrupt his competition." The statement follows a similar comment made by Sam Altman, as tensions between Musk and the OpenAI and Altman duo continue to grow.
Perplexity has joined the AI research race with OpenAI with the release of its own new freemium Deep Research product, designed to deliver detailed, citation-backed reports for expert-level tasks in areas like finance, marketing, and product research. Currently available on the web and shortly on Mac, iOS, and Android apps, the tool "iteratively searches, reads documents, and reasons about what to do next," which mimics the research process of most people. Currently, Perplexity is offering a free tier with limited queries and outpaces competitors by completing most tasks in just under three minutes.
South Korea has hit the breaks on new downloads of China's DeepSeek AI chatbot, claiming the model needs "improvements and remedies" to comply with the country's personal data protection laws. Enforced by the Personal Information Protection Commission, the ban follows DeepSeek's rapid rise after reaching over a million weekly users and causing subsequent privacy and national security concerns. While existing users will have no problems continuing to access the app or use its website, government agencies have already prohibited downloads on work devices.
The New York Times is now authorizing AI tools - like its new internal tool Echo, GitHub Copilot, and Google Vertex AI - to assist with tasks such as editing, summarizing, coding, and generating promotional content. The move comes with clear guidelines ensuring that AI will only ever serve as a support tool, with all AI-made outputs vetted and reviewed by expert journalists. All these tools and guidelines are rolling out in the middle of The New York Times' legal scuffle with OpenAI and Microsoft, where the publisher alleges that ChatGPT was trained on its content.
Dream, an AI-based cybersecurity company focused on critical national infrastructure, recently announced in a press release that it had secured $100 million in Series B funding, putting its valuation at $1.1 billion. Former Austrian prime minister and co-founder Sebastian Kurz highlighted the real human impact of cyberattacks, emphasizing that Dream's AI models "think both like an attacker and a defender" to pre-empt threats. The company has reported over $130 million in 2024 sales, making it a strong competitor to expand into new markets.
|
Nia - Nia is a context-aware collaborative teammate that understands your codebase for instant architecture insights, multi-file references, automated code reviews, and more.
Smodin - Made for students, teachers, and writers, Smodin is an AI writing tool for writing essays, articles, and paraphrasing text with features such as rewriters, summarizers, plagiarism checkers, and multi-lingual translators.
Shortwave - Organize your inbox the way you want - Shortwave helps you streamline email management with its AI assistant by analyzing your email history to help you write content drafts and pick out what's important.
Echo - Think deeper with Echo, an AI-powered voice and text note-taking app that refines your thoughts and brings your notes to life with smart organization.
|
Wake County's AI Playbook: Leading the AI Revolution (4‑min read)
North Carolina's Wake County is crafting an AI playbook to harness transformative tech: leveraging strategic partnerships, workforce training, and ethical guidelines, here's how it aims to become a leader in US AI by February 2025.
|
| What did you think of this newsletter? Let us know! |
|
|
|
|
|