AI

Gemini’s data-analyzing abilities aren’t as good as Google claims

Comment

In this photo illustration a Gemini logo and a welcome message on Gemini website are displayed on two screens.
Image Credits: Lorenzo Di Cola/NurPhoto / Getty Images

One of the selling points of Google’s flagship generative AI models, Gemini 1.5 Pro and 1.5 Flash, is the amount of data they can supposedly process and analyze. In press briefings and demos, Google has repeatedly claimed that the models can accomplish previously impossible tasks thanks to their “long context,” like summarizing multiple hundred-page documents or searching across scenes in film footage.

But new research suggests that the models aren’t, in fact, very good at those things.

Two separate studies investigated how well Google’s Gemini models and others make sense out of an enormous amount of data — think “War and Peace”-length works. Both find that Gemini 1.5 Pro and 1.5 Flash struggle to answer questions about large datasets correctly; in one series of document-based tests, the models gave the right answer only 40%-50% of the time.

“While models like Gemini 1.5 Pro can technically process long contexts, we have seen many cases indicating that the models don’t actually ‘understand’ the content,” Marzena Karpinska, a postdoc at UMass Amherst and a co-author on one of the studies, told TechCrunch.

Gemini’s context window is lacking

A model’s context, or context window, refers to input data (e.g., text) that the model considers before generating output (e.g., additional text). A simple question — “Who won the 2020 U.S. presidential election?” — can serve as context, as can a movie script, show or audio clip. And as context windows grow, so does the size of the documents being fit into them.

The newest versions of Gemini can take in upward of 2 million tokens as context. (“Tokens” are subdivided bits of raw data, like the syllables “fan,” “tas” and “tic” in the word “fantastic.”) That’s equivalent to roughly 1.4 million words, two hours of video or 22 hours of audio — the largest context of any commercially available model.

In a briefing earlier this year, Google showed several pre-recorded demos meant to illustrate the potential of Gemini’s long-context capabilities. One had Gemini 1.5 Pro search the transcript of the Apollo 11 moon landing telecast — around 402 pages — for quotes containing jokes, and then find a scene in the telecast that looked similar to a pencil sketch.

VP of research at Google DeepMind Oriol Vinyals, who led the briefing, described the model as “magical.”

“[1.5 Pro] performs these sorts of reasoning tasks across every single page, every single word,” he said.

That might have been an exaggeration.

In one of the aforementioned studies benchmarking these capabilities, Karpinska, along with researchers from the Allen Institute for AI and Princeton, asked the models to evaluate true/false statements about fiction books written in English. The researchers chose recent works so that the models couldn’t “cheat” by relying on foreknowledge, and they peppered the statements with references to specific details and plot points that’d be impossible to comprehend without reading the books in their entirety.

Given a statement like “By using her skills as an Apoth, Nusis is able to reverse engineer the type of portal opened by the reagents key found in Rona’s wooden chest,” Gemini 1.5 Pro and 1.5 Flash — having ingested the relevant book — had to say whether the statement was true or false and explain their reasoning.

Image Credits: UMass Amherst

Tested on one book around 260,000 words (~520 pages) in length, the researchers found that 1.5 Pro answered the true/false statements correctly 46.7% of the time while Flash answered correctly only 20% of the time. That means a coin is significantly better at answering questions about the book than Google’s latest machine learning model. Averaging all the benchmark results, neither model managed to achieve higher than random chance in terms of question-answering accuracy.

“We’ve noticed that the models have more difficulty verifying claims that require considering larger portions of the book, or even the entire book, compared to claims that can be solved by retrieving sentence-level evidence,” Karpinska said. “Qualitatively, we also observed that the models struggle with verifying claims about implicit information that is clear to a human reader but not explicitly stated in the text.”

The second of the two studies, co-authored by researchers at UC Santa Barbara, tested the ability of Gemini 1.5 Flash (but not 1.5 Pro) to “reason over” videos — that is, search through and answer questions about the content in them.

The co-authors created a dataset of images (e.g., a photo of a birthday cake) paired with questions for the model to answer about the objects depicted in the images (e.g., “What cartoon character is on this cake?”). To evaluate the models, they picked one of the images at random and inserted “distractor” images before and after it to create slideshow-like footage.

Flash didn’t perform all that well. In a test that had the model transcribe six handwritten digits from a “slideshow” of 25 images, Flash got around 50% of the transcriptions right. The accuracy dropped to around 30% with eight digits.

“On real question-answering tasks over images, it appears to be particularly hard for all the models we tested,” Michael Saxon, a PhD student at UC Santa Barbara and one of the study’s co-authors, told TechCrunch. “That small amount of reasoning — recognizing that a number is in a frame and reading it — might be what is breaking the model.”

Google is overpromising with Gemini

Neither of the studies have been peer-reviewed, nor do they probe the releases of Gemini 1.5 Pro and 1.5 Flash with 2-million-token contexts. (Both tested the 1-million-token context releases.) And Flash isn’t meant to be as capable as Pro in terms of performance; Google advertises it as a low-cost alternative.

Nevertheless, both add fuel to the fire that Google’s been overpromising — and under-delivering — with Gemini from the beginning. None of the models the researchers tested, including OpenAI’s GPT-4o and Anthropic’s Claude 3.5 Sonnet, performed well. But Google’s the only model provider that’s given context window top billing in its advertisements.

“There’s nothing wrong with the simple claim, ‘Our model can take X number of tokens’ based on the objective technical details,” Saxon said. “But the question is, what useful thing can you do with it?”

Generative AI broadly speaking is coming under increased scrutiny as businesses (and investors) grow frustrated with the technology’s limitations.

In a pair of recent surveys from Boston Consulting Group, about half of the respondents — all C-suite executives — said that they don’t expect generative AI to bring about substantial productivity gains and that they’re worried about the potential for mistakes and data compromises arising from generative AI-powered tools. PitchBook recently reported that, for two consecutive quarters, generative AI dealmaking at the earliest stages has declined, plummeting 76% from its Q3 2023 peak.

Faced with meeting-summarizing chatbots that conjure up fictional details about people and AI search platforms that basically amount to plagiarism generators, customers are on the hunt for promising differentiators. Google — which has raced, at times clumsily, to catch up to its generative AI rivals — was desperate to make Gemini’s context one of those differentiators.

But the bet was premature, it seems.

“We haven’t settled on a way to really show that ‘reasoning’ or ‘understanding’ over long documents is taking place, and basically every group releasing these models is cobbling together their own ad hoc evals to make these claims,” Karpinska said. “Without the knowledge of how long context processing is implemented — and companies do not share these details — it is hard to say how realistic these claims are.”

Google didn’t respond to a request for comment.

Both Saxon and Karpinska believe the antidotes to hyped-up claims around generative AI are better benchmarks and, along the same vein, greater emphasis on third-party critique. Saxon notes that one of the more common tests for long context (liberally cited by Google in its marketing materials), “needle in the haystack,” only measures a model’s ability to retrieve particular info, like names and numbers, from datasets — not answer complex questions about that info.

“All scientists and most engineers using these models are essentially in agreement that our existing benchmark culture is broken,” Saxon said, “so it’s important that the public understands to take these giant reports containing numbers like ‘general intelligence across benchmarks’ with a massive grain of salt.”

More TechCrunch

Anthropic is launching a program to fund the development of new types of benchmarks capable of evaluating the performance and impact of AI models, including generative models like its own…

Anthropic looks to fund a new, more comprehensive generation of AI benchmarks

A group of senators has banded together to urge Synapse’s owners and bank and fintech partners to “immediately restore customers’ access to their money.” As part of their demands, the…

Senators urge owners, partners, and VC backers of fintech Synapse to restore customers’ access to their money

Hello and welcome back to TechCrunch Space. I hope everyone has a fantastic July 4 this week. Go eat a hot dog. Read my story from last week on the…

TechCrunch Space: Star spangled

Music, podcasts, audiobooks…emergency alerts? Spotify’s latest test has the streaming app venturing into new territory with a test of an emergency alerts system in its home market of Sweden. According…

Spotify tests emergency alerts in Sweden

Simply submitting the request for a takedown doesn’t necessarily mean the content will be removed, however.

YouTube now lets you request removal of AI-generated content that simulates your face or voice

The news highlights that the fallout from the Evolve data breach on third-party companies — and their customers and users —  is still unclear.

Fintech company Wise says some customers affected by Evolve Bank data breach

The Supreme Court on Monday vacated two judicial decisions concerning Republican-backed laws from Florida and Texas aimed at limiting social media companies’ ability to moderate content on their platforms. The…

Supreme Court sends Texas and Florida social media regulation laws back to lower courts

Afloat, a gift delivery app that lets you shop from local stores and have gifts delivered to a loved one on the same day, is now available across the U.S. The…

Gifting on-demand startup Afloat goes nationwide

Exciting news for tech enthusiasts and innovators! TechCrunch Disrupt 2024 is just around the corner, and we have an incredible opportunity for you to elevate your brand’s visibility. How? By…

Drive brand impact with a Side Event at TechCrunch Disrupt

After Meta started tagging photos with a “Made with AI” label in May, photographers complained that the social networking company had been applying labels to real photos where they had…

Meta changes its label from ‘Made with AI’ to ‘AI info’ to indicate use of AI in photos

Investment app Robinhood is adding more AI features for investors with its acquisition of AI-powered research platform Pluto Capital, Inc. Announced on Monday, the company says that Pluto will allow…

Robinhood snaps up Pluto to add AI tools to its investing app

Vaire Computing, based in London and Seattle, is betting that chips that can do reversible computing are going to be the way forward for the world.

Vaire Computing raises $4.5M for ‘reversible computing’ moonshot which could drastically reduce energy needs

The EC has found that Meta’s “pay or consent” offer to Facebook and Instagram users in Europe does not comply with the bloc’s DMA.

Meta’s ‘pay or consent’ model fails EU competition rules, Commission finds

The round was led by KKR and Teachers’ Ventures Growth, an investment arm of Ontario Teachers’ Pension Plan.

Japan’s SmartHR raises $140M Series E as strong demand for HR tech boosts its ARR to $100M

RoboGrocery combines computer vision with a soft robotic gripper to bag a wide range of different items.

MIT’s soft robotic system is designed to pack groceries

This is by no means a complete list, just a few of the most obvious tricks that AI can supercharge.

AI-powered scams and what you can do about them

Identity.vc writes checks that range from €250,000 to €1.5 million into companies from the pre-seed to Series A stages.

Identity.vc is bringing capital and community to Europe’s LGBTQ+ venture ecosystem

Featured Article

Robot cats, dogs and birds are being deployed amid an ‘epidemic of loneliness’

In the early 1990s, a researcher at Japan’s National Institute of Advanced Industrial Science and Technology began work on what would become Paro. More than 30 years after its development, the doe-eyed seal pup remains the best-known example of a therapeutic robot for older adults. In 2011, the robot reached…

1 day ago
Robot cats, dogs and birds are being deployed amid an ‘epidemic of loneliness’

Apple’s AI plans go beyond the previously announced Apple Intelligence launches on the iPhone, iPad and Mac. According to Bloomberg’s Mark Gurman, the company is also working to bring these…

Apple reportedly working to bring AI to the Vision Pro

One of the earlier SaaS adherents to generative AI has been ServiceNow, which has been able to take advantage of the data in its own platform to help build more…

ServiceNow’s generative AI solutions are taking advantage of the data on its own platform

India’s top AI startups include those building LLMs and setting up the stage for AGI as well as bringing AI to cooking and serving farmers.

Here are India’s biggest AI startups based on how much money they’ve raised

We live in a very different world since the Russian invasion of Ukraine in 2022 and Hamas’s October 7 attack on Israel. With global military expenditure reaching $2.4 trillion last…

Defense tech and ‘resilience’ get global funding sources: Here are some top funders

Two separate studies investigated how well Google’s Gemini models and others make sense out of an enormous amount of data.

Gemini’s data-analyzing abilities aren’t as good as Google claims
Image Credits: Lorenzo Di Cola/NurPhoto / Getty Images

Featured Article

The biggest data breaches in 2024: 1 billion stolen records and rising

Some of the largest, most damaging breaches of 2024 already account for over a billion stolen records.

2 days ago
The biggest data breaches in 2024: 1 billion stolen records and rising

Welcome back to TechCrunch’s Week in Review — TechCrunch’s newsletter recapping the week’s biggest news. Want it in your inbox every Saturday? Sign up here. This week, Apple finally added…

Apple finally supports RCS in iOS 18 update

Featured Article

SAP, and Oracle, and IBM, oh my! ‘Cloud and AI’ drive legacy software firms to record valuations

There’s something of a trend around legacy software firms and their soaring valuations: Companies founded in dinosaur times are on a tear, evidenced this week with SAP‘s shares topping $200 for the first time. Founded in 1972, SAP’s valuation currently sits at an all-time high of $234 billion. The Germany-based…

2 days ago
SAP, and Oracle, and IBM, oh my! ‘Cloud and AI’ drive legacy software firms to record valuations

Sarah Bitamazire is the chief policy officer at the boutique advisory firm Lumiera.

Women in AI: Sarah Bitamazire helps companies implement responsible AI

Crypto platforms will need to report transactions to the Internal Revenue Service, starting in 2026. However, decentralized platforms that don’t hold assets themselves will be exempt. Those are the main…

IRS finalizes new regulations for crypto tax reporting

As part of a legal settlement, the Detroit Police Department has agreed to new guardrails limiting how it can use facial recognition technology. These new policies prohibit the police from…

Detroit Police Department agrees to new rules around facial recognition tech

Plaid’s expansion into being a multi-product company has led to real traction beyond traditional fintech customers.

Plaid, once aimed at mostly fintechs, is growing its enterprise business and now has over 1,000 customers signed on
  翻译: