☕️ Did OpenAI's Fabled GPT-4.5 "Orion" Fall Short of Expectations?

With the industry eagerly waiting for the debut of OpenAI's "Orion" nearly a year after the last of its series, did GPT-4.5 end up delivering? Other key highlights include:

• Tencent's Hunyuan Turbo S model aims to be one of the fastest in the Chinese AI market

• Amazon releases a "next-generation Alexa" called Alexa+, jam-packed with AI features

• Reports of Meta's own AI chatbot app start circulating as LlamaCon draws closer

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

Why I Believe in Open-Source AI - And Why You Should Too!

"Proprietary LLMs are powerful but come with risks - cost, control, and compliance. That’s why I'm co-hosting this event to explore IBM's Granite, an enterprise-grade open-source LLM that gives you freedom to build AI on your terms."

- Mark Hinkle, your AI Sherpa

INTERESTED? CLICK HERE TO SIGN UP!

OpenAI releases GPT-4.5 "Orion," the last of its kind non-reasoning model

OpenAI's "Orion" finally sees the light of day with the release of GPT-4.5, just days after Anthropic got into the spotlight with Claude 3.7 Sonnet's public debut. GPT-4.5 is OpenAI's largest model yet trained with more computing power and data than any of the company's previous releases. As the final member of OpenAI's non-reasoning line of models, expectations were high, but it may not be all sunshine and rainbows after all.

Let's elaborate a little.

​The white paper on GPT-4.5 says it builds upon the simple traditional technique of drastically increasing the data and computing power fed into the model during the pre-training phase. Unfortunately, the returns of this method have begun to plateau, and although GPT-4.5 does fairly consistently win over GPT-4o in benchmarks, it struggles to consistently hold up against reasoning models from Anthropic, DeepSeek, or its own. Additionally, due to its massive size, GPT-4.5 is ridiculously expensive to run - $75/million input tokens and $150/million output tokens compared to GPT-4o's $2.50/million input tokens and $15/million output tokens.

There are some upsides, however - OpenAI reports that GPT-4.5 has the highest EQ (a measure of emotional intelligence) and considers it its "best model for chat." It is the best of OpenAI's models at "understanding human intent" and is one of the least likely models to "hallucinate," aka make things up on the fly, as measured on SimpleQA.

Tencent released a new AI model that replies faster than "slow thinking models" like DeepSeek-R1

China continues to deliver model after model as Tencent is next in line with the release of its Hunyuan Turbo S that distinguishes itself from the likes of DeepSeek's R1 with its one-second-fast replies. When tested on knowledge fields like math and reasoning, Tencent claims that the Turbo S squares off evenly against DeepSeek-V3. In addition to its speed, the deployment and usage costs of the model have been starkly reduced compared to previous iterations, which can be likely attributed to DeepSeek's low-pricing strategy.

​

Amazon debuts a new version of Alexa packed with generative AI capabilities

In the middle of a calm Wednesday, Amazon unveiled what it calls its next-generation digital voice assistant, aptly named Alexa+, a GenAI-enhanced version of the original that works with older Echo devices. Amazon plans for Alexa+ to launch in late March at at a $19.99/month price tag(free for Prime customers) and is designed to perform tasks like ordering groceries, analyzing documents, and even navigating the web. It leverages multiple AI models, including Amazon's Nova and Anthropic's models, for better flexibility in interpreting user requests.

​

Meta could be readying itself to launch its own standalone AI chatbot app as soon as April

​According to reports by CNBC, Meta could be gearing up to launch a standalone app for its AI assistant, Meta AI, to compete on an equal platform with rivals like ChatGPT and Gemini. CNBC reports the new app could debut as early as Meta's next fiscal quarter between April and June while also exploring an additional paid subscription service - no price figures publicly known. With over 700 million active monthly users, Meta AI continues to be a key part of the company's strategy, and it doesn't look to be slowing down with its April-bound AI-centric dev conference called LlamaCon.

​

Nvidia remains confident amid concerns over DeepSeek following strong demand for its AI chips

Nvidia rides off into the week following a strong quarterly earnings report, with AI chip sale revenues exceeding $39 billion, up 74% year-on-year, and shrugging off concerns from the surprise shock by DeepSeek. The company remains confident in its position as big tech firms continue to rely on its advanced chips to train AI models. Off the back of its Blackwell chip release, Nvidia is rapidly scaling production, and its shares have surged over 400% in the past two years as a result of its efforts.

​

Framework releases its first AI-ready inference and gaming PC packed with 4.5L of AMD power

Taking a step to the side from its usual repairable and modular laptop shtick, Framework is releasing its first-ever desktop PC. Built around AMD's new Strix Halo architecture, the compact 4.5L PC (smaller than a PS5 and Xbox Series X) is optimized for both gaming and local AI inference. The Framework Desktop comes with customizable features, soldered LPDDR5x memory for massive bandwidth, and support for both Windows and Linux. The PC is on offer in 3 different configurations, ranging from $1,099 to $1,999, with preorders open and shipments set for early Q3 2025.

​

Perplexity launches a $50 million US venture seed fund following recent rapid expansion

AI search engine platform Perplexity is reportedly in the middle of creating a $50 million venture fund focused on US-based pre-seed and seed AI startups to serve as an anchor investor, according to CNBC. This comes on the heels of a $500 million funding round that tripled its valuation to $9 billion in December and a small list of new products. Among its most recent announcements are its upcoming web browser, Comet, and an AI-powered shopping assistant, Buy With Pro, while its search engine handles over 100 million queries per week.

Your AI, Your Rules - Deploy Open-Source LLMs at Scale

"Open-source LLMs like IBM’s Granite offer the flexibility and transparency businesses need for real AI innovation. I’m co-hosting this talk to show how enterprises can build AI solutions without giving up control."

- Mark Hinkle, your AI Sherpa

​Caramel - Turn your ideas into campaigns in seconds with Caramel, an AI-powered ad creation and optimization platform for Facebook, Google, Instagram, and more without marketing experience.

​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.

​Parafact - Factcheck any human or AI-written text in real-time with Parafact, an AI-powered tool that tracks down reliable sources and citations - all in one click.

Amazon's $25 Billion Bet on Robotics (3-min read)

In an effort to slash costs with robots and drive a surge in fueling the worldwide AI boom, Amazon plans to invest up to $25 billion in warehouse automation as rivals like China's Temu and Shein continue to be a thorn up its side.