Tag: large language models

  • The World Runs on Excel… and its Mistakes

    The 2022 CHIPS and Science Act earmarked hundreds of billions in subsidies and tax credits to bolster a U.S. domestic semiconductor (and especially semiconductor manufacturing) industry. If it works, it will dramatically reposition the U.S. in the global semiconductor value chain (especially relative to China).

    With such large amounts of taxpayer money practically “gifted” to large (already very profitable) corporations like Intel, the U.S. taxpayer can reasonably assume that these funds should be allocated carefully and thoughtfully and with processes in place to make sure every penny furthered the U.S.’s strategic goals.

    But, when the world’s financial decisions are powered by Excel spreadsheets, even the best laid plans can go awry.

    The team behind the startup Rowsie created a large language model (LLM)-powered tool which can understand Excel spreadsheets and answer questions posed to it. They downloaded a spreadsheet that the US government provided as an example of the information and calculations they want applicants fill out in order to qualify. They then applied their AI tool to the spreadsheet to understand it’s structure and formulas.

    Interestingly, Rowsie was able to find a single-cell spreadsheet error (see images below) which resulted in a $178 million understatement of interest payments!

    The Control Panel tab in the Example Pre-App-Simple-Financial-Model spreadsheet from the CHIPS Act funding application website. Notice cells C51 and D51 corresponding to the senior debt interest rate (5%) and the subordinated debt interest rate (8%)
    The Assumptions Processing tab in the Example Pre-App-Simple-Financial-Model spreadsheet from the CHIPS Act funding application website. Notice row 50. Despite the section being about Subordinated Debt (see Cell B50), they’re using cell C51 from the Control Panel tab (which points to the Senior Debt rate of 5%) rather than the correct cell of D51 (which points to the Subordinated Debt rate of 8%).

    To be clear, this is not a criticism of the spreadsheet’s architects. In this case, what seems to have happened, is that the spreadsheet creator copied an earlier row (row 40) and forgot to edit the formula to account for the fact that row 50 is about subordinated debt and row 40 is about senior debt. It’s a familiar story to anyone who’s ever been tasked with doing something complicated in Excel. Features like copy and paste and complex formulas are very powerful, but also make it very easy for a small mistake to cascade. It’s also remarkably hard to catch!

    Hopefully the Department of Commerce catches on and fixes this little clerical mishap, and that applicants are submitting good spreadsheets, free of errors. But, this case underscores how (1) so many of the world’s financial and policy decisions rest on Excel spreadsheets and you just have to hope 🤞🏻 no large mistakes were made, and (2) the potential for tools like Rowsie to be tireless proofreaders and assistants who can help us avoid mistakes and understand those critical spreadsheets quickly.

    If you’re interested in checking out Rowsie, check it out at https://www.rowsie.ai/!

    DISCLAIMER: I happen to be friends with the founders of Rowsie which is how I found out about this

  • The “Large Vision Model” (LVM) Era is Upon Us

    Unless you’ve been under a rock, you’ll know the tech industry has been rocked by the rapid advance in performance by large language models (LLMs) such as ChatGPT. By adapting self-supervised learning methods, LLMs “learn” to sound like a human being by learning how to fill in gaps in language and, by doing so, become remarkably adept at solving not just language problems but understanding & creativity.

    Interestingly, the same is happening in imaging, as models largely trained to fill in “gaps” in images are becoming amazingly adept. A friend of mine, Pearse Keane’s group at University College of London, for instance, just published a model trained using self-supervised learning methods on ophthalmological images which is capable of not only diagnosing diabetic retinopathy and glaucoma relatively accurately, but relatively good at predicting cardiovascular events and Parkinson’s.

    At a talk, Andrew Ng captured it well, by pointing out the parallels between the advances in language modeling that happened after the seminal Transformer paper and what is happening in the “large vision model” world with this great illustration.

    From Andrew Ng (Image credit: EETimes)

    “The revolution we’ve seen for text will be coming to images,” renowned computer scientist Andrew Ng asserted in a keynote talk he gave at the recent AI Hardware Summit here.

    Andrew Ng: The AI Text Revolution
    Is Coming to Images
    Sally Ward-Foxton | EETimes