AI

Meet two open source challengers to OpenAI’s ‘multimodal’ GPT-4V

Comment

Stable Diffusion
Image Credits: Bryce Durbin / TechCrunch

OpenAI’s GPT-4V is being hailed as the next big thing in AI: a “multimodal” model that can understand both text and images. This has obvious utility, which is why a pair of open source projects have released similar models — but there’s also a dark side that they may have more trouble handling. Here’s how they stack up.

Multimodal models can do things that strictly text- or image-analyzing models can’t. For instance, GPT-4V could provide instructions that are easier to show than tell, like fixing a bicycle. And because multimodal models can not only identify what’s in an image but extrapolate and comprehend the contents (at least to a degree), they go beyond the obvious — for example, suggesting recipes that can be prepared using ingredients from a pictured fridge.

But multimodal models present new risks. OpenAI initially held back the release of GPT-4V, fearing that it could be used to identify people in images without their consent or knowledge.

Even now, GPT-4V — which is only available to subscribers of OpenAI’s ChatGPT Plus plan — has worrisome flaws, including an inability to recognize hate symbols and a tendency to discriminate against certain sexes, demographics and body types. And this is according to OpenAI itself!

OpenAI’s GPT-4 with vision still has flaws, paper reveals

Open options

Despite the risks, companies — and loose cohorts of independent developers — are forging ahead, releasing open source multimodal models that, while not as capable as GPT-4V, can accomplish many, if not most, of the same things.

Earlier this month, a team of researchers from the University of Wisconsin-Madison, Microsoft Research and Columbia University released LLaVA-1.5 (an acronym for “Large Language-and-Vision Assistant”), which, like GPT-4V, can answer questions about images given prompts like “What’s unusual about this picture?” and “What are the things I should be cautious about when I visit here?”

LLaVA-1.5 followed on the heels of Qwen-VL, a multimodal model open sourced by a team at Alibaba (and which Alibaba is licensing to companies with over 100 million monthly active users), and image-and-text-understanding models from Google including PaLI-X and PaLM-E. But LLaVA-1.5 is one of the first multimodal models that’s easy to get up and running on consumer-level hardware — a GPU with less than 8GB of VRAM.

Elsewhere, Adept, a startup building AI models that can navigate software and the web autonomously, open sourced a GPT-4V-like multimodal text-and-image model — but with a twist. Adept’s model understands “knowledge worker” data such as charts, graphs and screens, enabling it to manipulate — and reason over — this data.

LLaVA-1.5

LLaVA-1.5 is an improved version of LLaVA, which was released several months ago by a Microsoft-affiliated research team.

Like LLaVA, LLaVA-1.5 combines a component called a “visual encoder” and Vicuna, an open source chatbot based on Meta’s Llama model, to make sense of images and text and how they relate.

The research team behind the original LLaVA generated the model’s training data using the text-only versions of OpenAI’s ChatGPT and GPT-4. They provided ChatGPT and GPT-4 with image descriptions and metadata, prompting the models to create conversations, questions, answers and reasoning problems based on the image content.

The LLaVA-1.5 team took this a step further by scaling up the image resolution and adding data including from ShareGPT, a platform where users share conversations with ChatGPT, to the LLaVA training dataset.

The larger of the two available LLaVA-1.5 models, which contains 13 billion parameters, can be trained in a day on eight Nvidia A100 GPUs, amounting to a few hundred dollars in server costs. (Parameters are the parts of a model learned from historical training data and essentially define the skill of the model on a problem, such as generating text.)

That’s not cheap, per se. But considering that it reportedly cost OpenAI tens of millions of dollars to train GPT-4, it’s definitely a step in the right direction. That is, if it performs well enough.

James Gallagher and Piotr Skalski, two software engineers at computer vision startup Roboflow, recently ran LLaVA-1.5 through its paces and detailed the results in a blog post.

First, they tested the model’s “zero-shot” object detection, or its ability to identify an object it wasn’t explicitly trained to recognize. They asked LLaVA-1.5 to detect a dog in an image, and, impressively, it managed to do this — even specifying the coordinates in the image where it “saw” the dog.

Llava-1.5
Image Credits: Roboflow

Gallagher and Skalski then moved on to a harder test: asking the model to explain a meme. Memes aren’t always easy for models (or even people) to understand, given their double meanings, entendres, in-jokes and subtext. So they make for a useful benchmark of a multimodal model’s ability to contextualize and analyze.

Gallagher and Skalski fed LLaVA-1.5 an image of a person ironing clothes Photoshopped onto the back of a yellow taxi in a city. They asked LLaVA-1.5 “What is unusual about this image?” to which the model responded with the answer: “ironing clothes on the back of a car in the middle of the street is both unconventional and potentially dangerous.” Tough to argue with that logic. 

Llava-1.5
Image Credits: Roboflow

It’s in Gallagher’s and Skalski’s next few tests that LLaVA-1.5’s weaknesses began to show.

While they found the model could successfully figure out a coin’s denomination from an image of a single coin, LLaVA-1.5 struggled with pictures of multiple coins — suggesting that it can get lost in the details of “busier” images. 

Llava-1.5
Image Credits: Roboflow

LLaVA-1.5 also couldn’t reliably recognize text, in contrast to GPT-4V. When Gallagher and Skalski gave LLaVA-1.5 a screenshot of text from a web page, LLaVA-1.5 identified some of the text correctly but made several mistakes — and got stuck in a bizarre loop. GPT-4V had no such issues. 

Llava-1.5
Image Credits: Roboflow

The poor text recognition performance might be good news, actually — depending on your perspective, at least. Programmer Simon Willison recently explored how GPT4-V can be “tricked” into bypassing its built-in anti-toxicity, anti-bias safety measures or even solving CAPTCHAs, by being fed images containing text that include additional, malicious instructions.

Were LLaVA-1.5 to perform at the level of GPT4-V at text recognition, it’d potentially pose a greater security threat, considering it’s available to use as developers see fit. 

Well, mostly as developers see fit. As it was trained on data generated by ChatGPT, LLaVA-1.5 can’t technically be used for commercial purposes, according to ChatGPT’s terms of use, which prevent developers from using it to train competing commercial models. Whether that stops anyone remains to be seen.

On the earlier subject of safety measures, in my own quick test, it quickly became apparent that LLaVA-1.5 isn’t bound by the same toxicity filters as GPT-4V.

Asked to give advice to a pictured larger women, LLaVA-1.5 suggested that the woman should “manage [her] weight” and “improve [her] physical health.” GPT-4V outright refused to answer.

Llava-1.5
LLaVA-1.5 implies that the pictured person is unhealthy purely based on their appearance. Image Credits: Roboflow

Adept

With its first open source multimodal model, Fuyu-8B, Adept isn’t trying to compete with LLaVA-1.5. Like LLaVA-1.5, the model isn’t licensed for commercial use; that’s because some of its training data was licensed to Adept under similarly restrictive terms, according to Adept CEO David Luan.

Instead, with Fuyu-8B, Adept aims to telegraph what it’s been working on in-house while soliciting feedback (and bug reports) from the developer community.

“Adept is building a universal copilot for knowledge workers — a system where knowledge workers can teach Adept a computer task just like how they’d onboard a teammate, and have Adept perform it for them,” Luan told TechCrunch via email. “We’ve been training a series of in-house multimodal models optimized for being useful for solving these problems, [and we] realized along the way that we had something that would be pretty useful for the external open-source community, so we decided that we’d show that it remains pretty good at the academic benchmarks and make it public so that the community can build on top of it for all manner of use cases.”

Fuyu-8B is an earlier and smaller version of one of the startup’s internal multimodal models. Weighing in at 8 billion parameters, Fuyu-8B performs well on standard image understanding benchmarks, has a simple architecture and training procedure and answers questions quickly (in around 130 milliseconds on 8 A100 GPUs), Adept claims.

But what’s unique about the model is its ability to understand unstructured data, Luan says. Unlike LLaVA-1.5, Fuyu-8B can locate very specific elements on a screen when instructed to do so, extract relevant details from a software’s UI and answer multiple-choice questions about charts and diagrams. 

Or rather, it can theoretically. Fuyu-8B doesn’t come with these capabilities built in. Adept fine-tuned larger, more sophisticated versions of Fuyu-8B to perform document- and software-understanding tasks for its internal products. 

“Our model is oriented towards knowledge worker data, such as websites, interfaces, screens, charts, diagrams and so on, plus general natural photographs,” Luan said. “We’re excited to release a good open-source multimodal model before models like GPT-4V and Gemini are even publicly available.”

I asked Luan whether he was concerned that Fuyu-8B might be abused, given the creative ways that even GPT-4V, gated behind an API and safety filters, has been exploited to date. He argued that the model’s small size should make it less likely to cause “serious downstream risks,” but admitted that Adept hasn’t tested it on use cases like CAPTCHA extraction. 

“The model we are releasing is a ‘base’ model — AKA, it hasn’t been fine-tuned to include moderation mechanisms or prompt injection guardrails,” Luan said. “Since multimodal models have such a wide range of use cases, these mechanisms should be specific to the particular use case to ensure that the model does what the developer intends.”

Is that the wisest choice? I’m not so sure. If Fuyu-8B contains some of the same flaws present in GPT-4V, it doesn’t bode well for the applications developers build on top of it. Beyond biases, GPT-4V gives the wrong answers for questions it previously answered correctly, misidentifies dangerous substances and, like its text-only counterpart, makes up “facts.”

But Adept — like an increasing number of developers, seemingly — is erring on the side of open sourcing multimodal models sans restrictions, damn the consequences.

More TechCrunch

StrictlyVC events deliver exclusive insider content from the Silicon Valley & Global VC scene while creating meaningful connections over cocktails and canapés with leading investors, entrepreneurs and executives. And TechCrunch…

Meesho, a leading e-commerce startup in India, has secured $275 million in a new funding round.

Meesho, an Indian social commerce platform with 150M transacting users, raises $275M

Some Indian government websites have allowed scammers to plant advertisements capable of redirecting visitors to online betting platforms. TechCrunch discovered around four dozen “gov.in” website links associated with Indian states,…

Scammers found planting online betting ads on Indian government websites

Around 550 employees across autonomous vehicle company Motional have been laid off, according to information taken from WARN notice filings and sources at the company.  Earlier this week, TechCrunch reported…

Motional cut about 550 employees, around 40%, in recent restructuring, sources say

The deck included some redacted numbers, but there was still enough data to get a good picture.

Pitch Deck Teardown: Cloudsmith’s $15M Series A deck

The company is describing the event as “a chance to demo some ChatGPT and GPT-4 updates.”

OpenAI’s ChatGPT announcement: What we know so far

Unlike ChatGPT, Claude did not become a new App Store hit.

Anthropic’s Claude sees tepid reception on iOS compared with ChatGPT’s debut

Welcome to Startups Weekly — Haje‘s weekly recap of everything you can’t miss from the world of startups. Sign up here to get it in your inbox every Friday. Look,…

Startups Weekly: Trouble in EV land and Peloton is circling the drain

Scarcely five months after its founding, hard tech startup Layup Parts has landed a $9 million round of financing led by Founders Fund to transform composites manufacturing. Lux Capital and Haystack…

Founders Fund leads financing of composites startup Layup Parts

AI startup Anthropic is changing its policies to allow minors to use its generative AI systems — in certain circumstances, at least.  Announced in a post on the company’s official…

Anthropic now lets kids use its AI tech — within limits

Zeekr’s market hype is noteworthy and may indicate that investors see value in the high-quality, low-price offerings of Chinese automakers.

The buzziest EV IPO of the year is a Chinese automaker

Venture capital has been hit hard by souring macroeconomic conditions over the past few years and it’s not yet clear how the market downturn affected VC fund performance. But recent…

VC fund performance is down sharply — but it may have already hit its lowest point

The person who claims to have 49 million Dell customer records told TechCrunch that he brute-forced an online company portal and scraped customer data, including physical addresses, directly from Dell’s…

Threat actor says he scraped 49M Dell customer addresses before the company found out

The social network has announced an updated version of its app that lets you offer feedback about its algorithmic feed so you can better customize it.

Bluesky now lets you personalize main Discover feed using new controls

Microsoft will launch its own mobile game store in July, the company announced at the Bloomberg Technology Summit on Thursday. Xbox president Sarah Bond shared that the company plans to…

Microsoft is launching its mobile game store in July

Smart ring maker Oura is launching two new features focused on heart health, the company announced on Friday. The first claims to help users get an idea of their cardiovascular…

Oura launches two new heart health features

Keeping up with an industry as fast-moving as AI is a tall order. So until an AI can do it for you, here’s a handy roundup of recent stories in the world…

This Week in AI: OpenAI considers allowing AI porn

Garena is quietly developing new India-themed games even though Free Fire, its biggest title, has still not made a comeback to the country.

Garena is quietly making India-themed games even as Free Fire’s relaunch remains doubtful

The U.S.’ NHTSA has opened a fourth investigation into the Fisker Ocean SUV, spurred by multiple claims of “inadvertent Automatic Emergency Braking.”

Fisker Ocean faces fourth federal safety probe

CoreWeave has formally opened an office in London that will serve as its European headquarters and home to two new data centers.

CoreWeave, a $19B AI compute provider, opens European HQ in London with plans for 2 UK data centers

The Series C funding, which brings its total raise to around $95 million, will go toward mass production of the startup’s inaugural products

AI chip startup DEEPX secures $80M Series C at a $529M valuation 

A dust-up between Evolve Bank & Trust, Mercury and Synapse has led TabaPay to abandon its acquisition plans of troubled banking-as-a-service startup Synapse.

Infighting among fintech players has caused TabaPay to ‘pull out’ from buying bankrupt Synapse

The problem is not the media, but the message.

Apple’s ‘Crush’ ad is disgusting

The Twitter for Android client was “a demo app that Google had created and gave to us,” says Particle co-founder and ex-Twitter employee Sara Beykpour.

Google built some of the first social apps for Android, including Twitter and others

WhatsApp is updating its mobile apps for a fresh and more streamlined look, while also introducing a new “darker dark mode,” the company announced on Thursday. The messaging app says…

WhatsApp’s latest update streamlines navigation and adds a ‘darker dark mode’

Plinky lets you solve the problem of saving and organizing links from anywhere with a focus on simplicity and customization.

Plinky is an app for you to collect and organize links easily

The keynote kicks off at 10 a.m. PT on Tuesday and will offer glimpses into the latest versions of Android, Wear OS and Android TV.

Google I/O 2024: How to watch

For cancer patients, medicines administered in clinical trials can help save or extend lives. But despite thousands of trials in the United States each year, only 3% to 5% of…

Triomics raises $15M Series A to automate cancer clinical trials matching

Welcome back to TechCrunch Mobility — your central hub for news and insights on the future of transportation. Sign up here for free — just click TechCrunch Mobility! Tap, tap.…

Tesla drives Luminar lidar sales and Motional pauses robotaxi plans

The newly announced “Public Content Policy” will now join Reddit’s existing privacy policy and content policy to guide how Reddit’s data is being accessed and used by commercial entities and…

Reddit locks down its public data in new content policy, says use now requires a contract