AI Top-of-Mind for 7.22.24 — GPT-4o-mini

dave ginsburg
AI.society
Published in
4 min readJul 22, 2024

Today: GPT-4o-mini, OpenAI whistleblowers, more on SpreadsheetLLM and Project Strawberry, and Miss AI.

Top-of-mind is OpenAI and their new GPT-4o-mini, pushing the art on smaller models. ‘TechCrunch’ offers details, looking at speed and cost efficiencies, with comparisons to Gemini 1.5 Flash and Claude 3 Haiku. The model supports text and vision today, with audio and video planned.

Source: OpenAI

More OpenAI news, with three other items. The first, as posted by Will Lockett in ‘Predict’ looks at their NDAs and ongoing investigations. From the article:

Reuters recently reported that OpenAI whistleblowers have filed a complaint with the U.S. Securities and Exchange Commission, calling for an investigation over the company’s allegedly restrictive non-disclosure agreements (NDAs). Apparently, OpenAI has made their employees sign NDAs that waive their federal rights to whistleblower compensation!

And…

“Safety culture and processes have taken a backseat to shiny products”

Then, reporting by ‘The Information’ on OpenAI’s discussions with Broadcom about AI chip development.

· A new server chip that would rival the kind made by Nvidia is a long shot that would take years to come to fruition. And in trying to develop a chip, OpenAI risks upsetting Nvidia, OpenAI’s most important chip supplier.

· But it could also provide OpenAI with potential leverage in future pricing negotiations with the company. Nvidia has been generating unprecedented profit margins and sales on its AI-focused graphics processing units because customers such as OpenAI don’t have viable alternatives.

· Another key part of Altman’s plan involves building new data centers to house the chips. Recently, he told one industry executive that he aimed to set up one or more companies with outside investors to pay for real estate, power, data centers and the specialized AI chip servers that would be housed in them. He said OpenAI would commit to renting those servers, this person said.

Finally, an update by Ignacio de Gregorio on project ‘Strawberry.’ From his post that goes into the technical details:

· Importantly, understanding Project Strawberry gives insight into how the real moat in AI, reasoning, might be achieved; a dire need for OpenAI considering that current LLMs, which are heavily commoditized, no longer offer that promise.

· In a nutshell, instead of simply providing the first answer they predict, they iterate over their own ‘thoughts’ until they find the best one.

And a diagram that captures the essence of the research:

The LLM generates possible thought paths, and the verifier (another, smaller LLM) validates every thought, helping search for the best possible solution.

Source: Ignacio de Gregorio

Have you wondered how spreadsheets can be wedded with LLMs? Well, Ali Waseem details Microsoft’s new ‘SpreadsheetLLM’ that transforms content into an LLM-friendly format. And Ignacio de Gregorio offersadditional technical details, with the following:

· For reference, a spreadsheet with 576 rows and 23 columns yields 61k tokens, almost half of GPT-4o’s entire context window.

· This leaves little room for the user’s query and any additional context the model might require, such as information from other sources.

Read on for how Microsoft intends to solve the above problem, with a link to the original paper. SpreadsheetLLM is still a research project but stay tuned.

Source: Microsoft

And finally, two off-the-wall AI stories. Pushing the boundaries of AI, and not necessarily in a good way, is the first ‘Miss AI’ beauty pageant. ‘Ars Technica’ reports on the contest’s background as well as expected controversy.

And the second is AI-assisted racing, as reported by ‘EE Times.’ From the article:

· How do you drive a car on a racetrack at up to 200 mph when you do not have hands or legs? Well, the answer is not an autonomous vehicle, but clever deployment of edge AI using an Nvidia module, stereo vision camera and fast inferencing in the car itself with minimal latency using custom trained models for head tracking.

· That is what the semi-autonomous motorcar (SAM) car is — a project developed by Arrow Electronics in which its engineers modified the steering system of a McLaren 720S Spyder using an Nvidia Jetson AGX Orin module. This allowed former IndyCar series racing car driver Sam Schmidt to take to the wheel again after a crash in 2000 left him paralyzed from the shoulders down.

Don’t try this at home!

--

--

dave ginsburg
AI.society

Lifelong technophile and author with background in networking, security, the cloud, IIoT, and AI. Father. Winemaker. Husband of @mariehattar.