Connect with us

Tech

Vera wants to use AI to cull generative models’ worst behaviors

Published

on

Liz O’Sullivan is on a mission to make AI “a little bit safer,” in her own words.

A member of the National AI Advisory Committee, which drafts recommendations to the White House and Congress on how to foster AI adoption while regulating its risks, O’Sullivan spent 12 years on the business side of AI startups overseeing data labeling and operations and customer success. In 2019, she took a job at the Surveillance Technology Oversight Project, mounting campaigns to protect New Yorkers’ civil liberties, and co-founded Arthur AI, a startup that partners with civil society and academy to shine light into AI’s “black box.”

Now, O’Sullivan is gearing up for her next act with Vera, a startup building a toolkit that allows companies to establish “acceptable use policies” for generative AI — the type of AI models that generate text, images, music and more — and enforce these policies across open source and custom models.

Vera today closed a $2.7 million funding round led by Differential Venture Partners with participation from Essence VC, Everywhere VC, Betaworks, Greycroft and ATP Ventures. Bringing Vera’s total raised to $3.3 million, the new cash will be put toward growing Vera’s five-person team, R&D and scaling enterprise deployments, O’Sullivan says.

“Vera was founded because we’ve seen, firsthand, the power of AI to address real problems, just as we’ve seen the wild and wacky ways it can cause damage to companies, the public and the world,” O’Sullivan told TechCrunch in an email interview. “We need to responsibly shepherd this technology into the world, and as companies race to define their generative AI strategies, we’re entering an age where it’s critical that we move beyond AI principles and into practice. Vera is a team that can actually help.

O’Sullivan co-founded Vera in 2021 with Justin Norman, formerly a research scientist at Cisco, a lead data scientist in Cloudera’s AI research lab and the VP of data science at Yelp. In September, Norman was appointed a member of the Department of the Navy Science and Technology board, which provides advice and counsel to the U.S. Navy on matters and policies relating to scientific, technical and related functions,

Vera’s platform attempts to identify risks in model inputs — for example, a prompt like “write a cover letter for a software engineering role” to a text-generating model — and block, redact or otherwise transform requests that might contain things like personally identifiable information, security credentials, intellectual property and prompt injection attacks. (Prompt injection attacks, essentially carefully-worded malicious prompts, are often used to “trick” models into bypassing safety filters.)

Vera also places constraints on what models can “say” in response to prompts, according to O’Sullivan, giving companies greater control over the behavior of their models in production.

How does Vera achieve this? By using what O’Sullivan describes as “proprietary language and vision models” that sit between users and internal or third-party models (e.g. OpenAI’s GPT-4) and detect problematic content. Vera can block “inappropriate” prompts to — or answers from a model in any form, O’Sullivan claims, whether text, code, image or video.

“Our deep tech approach to enforcing policies goes beyond passive forms of documentation and checklists to address the direct points at which these risks occur,” O’Sullivan said. “Our solution … prevents riskier responses that may include criminal material or encourage users to self-harm.”

Companies are certainly encountering challenges — mainly compliance-related — in adopting generative AI models for their purposes. They’re worried about their confidential data ending up with developers who trained the models on user data, for instance; in recent months, major corporations including Apple, Walmart and Verizon have banned employees from using tools like OpenAI’s ChatGPT.

And offensive models are obviously bad for publicity. No brand wants the text-generating model powering their customer service chatbot, say, to spout racial epithets or give self-destructive advice.

But this reporter wonders if Vera’s approach is as reliable as O’Sullivan suggests.

No model is perfect — not even Vera’s — and it’s been demonstrated time and time again that content moderation models are prone to a whole host of biases. Some AI models trained to detect toxicity in text see phrases in African-American Vernacular English, the informal grammar used by some Black Americans, as disproportionately “toxic.” Meanwhile, certain computer vision algorithms have been found to label thermometers held by Black people as “guns” while labeling thermometers held by light-skinned subjects as “electronic devices.”

To be fair to O’Sullivan, she doesn’t claim Vera’s models are bulletproof — only that they can cull the worst of a generative AI models’ behaviors. There may be some truth to that (depending on the model, at least) — and the degree to which Vera has iterated and refined its own models.

“Today’s AI hype cycle obscures the very serious, very present risks that affect humans alive today,” O’Sullivan said. “Where AI overpromises, we see real people hurt by unpredictable, harmful, toxic and potentially criminal model behavior … AI is a powerful tool and like any powerful tool, should be actively controlled so that its benefits outweigh these risks, which is why Vera exists.”

Vera’s possible shortcomings aside, the company has competition in the nascent market for model-moderating tech.

Similar to Vera, Nvidia’s NeMo Guardrails and Salesforce’s Einstein Trust Layer attempt to prevent text-generating models from retaining or regurgitating sensitive data, such as customer purchase orders and phone numbers. Microsoft provides an AI service to moderate text and image content, including from models. Elsewhere, startups like HiddenLayer, DynamoFL and Protect AI are creating tooling to defend generative AI models against prompt engineering attacks.

So far as I can tell, Vera’s value proposition is that it tackles a whole range of generative AI threats at once — or promises to at the very least. Assuming that the tech works as advertised, that’s bound to be attractive for companies in search of a one-stop content moderation, AI-model-attack-fighting shop.

Indeed, O’Sullivan says that Vera already has a handful of customers. The waitlist for more opens today.

“CTOs, CISOs and CIOs all over the world are struggling to strike the ideal balance between AI-enhanced productivity and the risks these models present,” O’Sullivan said. “Vera unlocks generative AI capabilities with policy enforcement that can be transferred not just to today’s models, but to future models without the vendor lock-in that occurs when you choose a one-model or one-size-fits-all approach to generative AI.”

Source link

Continue Reading
Click to comment

Leave a Reply

Your email address will not be published. Required fields are marked *

Tech

Miranda Bogen is creating solutions to help govern AI

Published

on

By


To give AI-focused women academics and others their well-deserved — and overdue — time in the spotlight, TechCrunch is launching a series of interviews focusing on remarkable women who’ve contributed to the AI revolution. We’ll publish several pieces throughout the year as the AI boom continues, highlighting key work that often goes unrecognized. Read more profiles here.

Miranda Bogen is the founding director of the Center of Democracy and Technology’s AI Governance Lab, where she works to help create solutions that can effectively regulate and govern AI systems. She helped guide responsible AI strategies at Meta and previously worked as a senior policy analyst at the organization Uptown, which seeks to use tech to advance equity and justice.

Briefly, how did you get your start in AI? What attracted you to the field?

I was drawn to work on machine learning and AI by seeing the way these technologies were colliding with fundamental conversations about society — values, rights, and which communities get left behind. My early work exploring the intersection of AI and civil rights reinforced for me that AI systems are far more than technical artifacts; they are systems that both shape and are shaped by their interaction with people, bureaucracies, and policies. I’ve always been adept at translating between technical and non-technical contexts, and I was energized by the opportunity to help break through the appearance of technical complexity to help communities with different kinds of expertise shape the way AI is built from the ground up.

What work are you most proud of (in the AI field)?

When I first started working in this space, many folks still needed to be convinced AI systems could result in discriminatory impact for marginalized populations, let alone that anything needed to be done about those harms. While there is still too wide a gap between the status quo and a future where biases and other harms are tackled systematically, I’m gratified that the research my collaborators and I conducted on discrimination in personalized online advertising and my work within the industry on algorithmic fairness helped lead to meaningful changes to Meta’s ad delivery system and progress toward reducing disparities in access to important economic opportunities.

How do you navigate the challenges of the male-dominated tech industry and, by extension, the male-dominated AI industry?

I’ve been lucky to work with phenomenal colleagues and teams who have been generous with both opportunities and sincere support, and we tried to bring that energy into any room we found ourselves in. In my most recent career transition, I was delighted that nearly all of my options involved working on teams or within organizations led by phenomenal women, and I hope the field continues to lift up the voices of those who haven’t traditionally been centered in technology-oriented conversations.

What advice would you give to women seeking to enter the AI field?

The same advice I give to anyone who asks: find supportive managers, advisors, and teams who energize and inspire you, who value your opinion and perspective, and who put themselves on the line to stand up for you and your work.

What are some of the most pressing issues facing AI as it evolves?

The impacts and harms AI systems are already having on people are well-known at this point, and one of the biggest pressing challenges is moving beyond describing the problem to developing robust approaches for systematically addressing those harms and incentivizing their adoption. We launched the AI Governance Lab at CDT to drive progress in both directions.

What are some issues AI users should be aware of?

For the most part, AI systems are still missing seat belts, airbags, and traffic signs, so proceed with caution before using them for consequential tasks.

What is the best way to responsibly build AI?

The best way to responsibly build AI is with humility. Consider how the success of the AI system you are working on has been defined, who that definition serves, and what context may be missing. Think about for whom the system might fail and what will happen if it does. And build systems not just with the people who will use them but with the communities who will be subject to them.

How can investors better push for responsible AI?

Investors need to create room for technology builders to move more deliberately before rushing half-baked technologies to market. Intense competitive pressure to release the newest, biggest, and shiniest new AI models is leading to concerning underinvestment in responsible practices. While uninhibited innovation sings a tempting siren song, it is a mirage that will leave everyone worse off.

AI is not magic; it’s just a mirror that is being held up to society. If we want it to reflect something different, we’ve got work to do.



Source link

Continue Reading

Tech

This Week in AI: Addressing racism in AI image generators

Published

on

By


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 of machine learning, along with notable research and experiments we didn’t cover on their own.

This week in AI, Google paused its AI chatbot Gemini’s ability to generate images of people after a segment of users complained about historical inaccuracies. Told to depict “a Roman legion,” for instance, Gemini would show an anachronistic, cartoonish group of racially diverse foot soldiers while rendering “Zulu warriors” as Black.

It appears that Google — like some other AI vendors, including OpenAI — had implemented clumsy hardcoding under the hood to attempt to “correct” for biases in its model. In response to prompts like “show me images of only women” or “show me images of only men,” Gemini would refuse, asserting such images could “contribute to the exclusion and marginalization of other genders.” Gemini was also loath to generate images of people identified solely by their race — e.g. “white people” or “black people” — out of ostensible concern for “reducing individuals to their physical characteristics.”

Right wingers have latched on to the bugs as evidence of a “woke” agenda being perpetuated by the tech elite. But it doesn’t take Occam’s razor to see the less nefarious truth: Google, burned by its tools’ biases before (see: classifying Black men as gorillas, mistaking thermal guns in Black people’s hands as weapons, etc.), is so desperate to avoid history repeating itself that it’s manifesting a less biased world in its image-generating models — however erroneous.

In her best-selling book “White Fragility,” anti-racist educator Robin DiAngelo writes about how the erasure of race — “color blindness,” by another phrase — contributes to systemic racial power imbalances rather than mitigating or alleviating them. By purporting to “not see color” or reinforcing the notion that simply acknowledging the struggle of people of other races is sufficient to label oneself “woke,” people perpetuate harm by avoiding any substantive conservation on the topic, DiAngelo says.

Google’s ginger treatment of race-based prompts in Gemini didn’t avoid the issue, per se — but disingenuously attempted to conceal the worst of the model’s biases. One could argue (and many have) that these biases shouldn’t be ignored or glossed over, but addressed in the broader context of the training data from which they arise — i.e. society on the world wide web.

Yes, the data sets used to train image generators generally contain more white people than Black people, and yes, the images of Black people in those data sets reinforce negative stereotypes. That’s why image generators sexualize certain women of color, depict white men in positions of authority and generally favor wealthy Western perspectives.

Some may argue that there’s no winning for AI vendors. Whether they tackle — or choose not to tackle — models’ biases, they’ll be criticized. And that’s true. But I posit that, either way, these models are lacking in explanation — packaged in a fashion that minimizes the ways in which their biases manifest.

Were AI vendors to address their models’ shortcomings head on, in humble and transparent language, it’d go a lot further than haphazard attempts at “fixing” what’s essentially unfixable bias. We all have bias, the truth is — and we don’t treat people the same as a result. Nor do the models we’re building. And we’d do well to acknowledge that.

Here are some other AI stories of note from the past few days:

  • Women in AI: TechCrunch launched a series highlighting notable women in the field of AI. Read the list here.
  • Stable Diffusion v3: Stability AI has announced Stable Diffusion 3, the latest and most powerful version of the company’s image-generating AI model, based on a new architecture.
  • Chrome gets GenAI: Google’s new Gemini-powered tool in Chrome allows users to rewrite existing text on the web — or generate something completely new.
  • Blacker than ChatGPT: Creative ad agency McKinney developed a quiz game, Are You Blacker than ChatGPT?, to shine a light on AI bias.
  • Calls for laws: Hundreds of AI luminaries signed a public letter earlier this week calling for anti-deepfake legislation in the U.S.
  • Match made in AI: OpenAI has a new customer in Match Group, the owner of apps including Hinge, Tinder and Match, whose employees will use OpenAI’s AI tech to accomplish work-related tasks.
  • DeepMind safety: DeepMind, Google’s AI research division, has formed a new org, AI Safety and Alignment, made up of existing teams working on AI safety but also broadened to encompass new, specialized cohorts of GenAI researchers and engineers.
  • Open models: Barely a week after launching the latest iteration of its Gemini models, Google released Gemma, a new family of lightweight open-weight models.
  • House task force: The U.S. House of Representatives has founded a task force on AI that — as Devin writes — feels like a punt after years of indecision that show no sign of ending.

More machine learnings

AI models seem to know a lot, but what do they actually know? Well, the answer is nothing. But if you phrase the question slightly differently… they do seem to have internalized some “meanings” that are similar to what humans know. Although no AI truly understands what a cat or a dog is, could it have some sense of similarity encoded in its embeddings of those two words that is different from, say, cat and bottle? Amazon researchers believe so.

Their research compared the “trajectories” of similar but distinct sentences, like “the dog barked at the burglar” and “the burglar caused the dog to bark,” with those of grammatically similar but different sentences, like “a cat sleeps all day” and “a girl jogs all afternoon.” They found that the ones humans would find similar were indeed internally treated as more similar despite being grammatically different, and vice versa for the grammatically similar ones. OK, I feel like this paragraph was a little confusing, but suffice it to say that the meanings encoded in LLMs appear to be more robust and sophisticated than expected, not totally naive.

Neural encoding is proving useful in prosthetic vision, Swiss researchers at EPFL have found. Artificial retinas and other ways of replacing parts of the human visual system generally have very limited resolution due to the limitations of microelectrode arrays. So no matter how detailed the image is coming in, it has to be transmitted at a very low fidelity. But there are different ways of downsampling, and this team found that machine learning does a great job at it.

Image Credits: EPFL

“We found that if we applied a learning-based approach, we got improved results in terms of optimized sensory encoding. But more surprising was that when we used an unconstrained neural network, it learned to mimic aspects of retinal processing on its own,” said Diego Ghezzi in a news release. It does perceptual compression, basically. They tested it on mouse retinas, so it isn’t just theoretical.

An interesting application of computer vision by Stanford researchers hints at a mystery in how children develop their drawing skills. The team solicited and analyzed 37,000 drawings by kids of various objects and animals, and also (based on kids’ responses) how recognizable each drawing was. Interestingly, it wasn’t just the inclusion of signature features like a rabbit’s ears that made drawings more recognizable by other kids.

“The kinds of features that lead drawings from older children to be recognizable don’t seem to be driven by just a single feature that all the older kids learn to include in their drawings. It’s something much more complex that these machine learning systems are picking up on,” said lead researcher Judith Fan.

Chemists (also at EPFL) found that LLMs are also surprisingly adept at helping out with their work after minimal training. It’s not just doing chemistry directly, but rather being fine-tuned on a body of work that chemists individually can’t possibly know all of. For instance, in thousands of papers there may be a few hundred statements about whether a high-entropy alloy is single or multiple phase (you don’t have to know what this means — they do). The system (based on GPT-3) can be trained on this type of yes/no question and answer, and soon is able to extrapolate from that.

It’s not some huge advance, just more evidence that LLMs are a useful tool in this sense. “The point is that this is as easy as doing a literature search, which works for many chemical problems,” said researcher Berend Smit. “Querying a foundational model might become a routine way to bootstrap a project.”

Last, a word of caution from Berkeley researchers, though now that I’m reading the post again I see EPFL was involved with this one too. Go Lausanne! The group found that imagery found via Google was much more likely to enforce gender stereotypes for certain jobs and words than text mentioning the same thing. And there were also just way more men present in both cases.

Not only that, but in an experiment, they found that people who viewed images rather than reading text when researching a role associated those roles with one gender more reliably, even days later. “This isn’t only about the frequency of gender bias online,” said researcher Douglas Guilbeault. “Part of the story here is that there’s something very sticky, very potent about images’ representation of people that text just doesn’t have.”

With stuff like the Google image generator diversity fracas going on, it’s easy to lose sight of the established and frequently verified fact that the source of data for many AI models shows serious bias, and this bias has a real effect on people.



Source link

Continue Reading

Tech

Humane pushes Ai Pin ship date to mid-April

Published

on

By


Hardware is difficult, to paraphrase a famous adage. First-generation products from new startups are notoriously so, regardless of how much money and excitement you’ve managed to drum up. Given all that, it’s likely few are too surprised that Humane’s upcoming Ai Pin has been pushed back a bit, from March to “mid-April,” per a new video from the Bay Area startup’s Head of Media, Sam Sheffer.

In the Sorkin-style walk and talk, he explains that the first units are set to, “start leaving the factory at the end of March.” If Humane keeps to that time frame, “priority access” customers will begin to receive the unit at some point in mid-April. The remaining preorders, meanwhile, should arrive “shortly after.”

Humane captured a good deal of tech buzz well before its first product was announced, courtesy of its founders’ time at Apple and some appropriately enigmatic prelaunch videos. The Ai Pin was finally unveiled at an event in San Francisco back in early November, where we were able to spend a little controlled hands-on time with the wearable.

The device is the first prominent example of what’s likely to be a growing trend in the consumer hardware world, as more startups look to harness the white-hot world of generative AI for new form factors. Humane is positioning its product as the next step for a space that’s been stuck on the smartphone form factor for more than a decade.

Image Credits: Humane

Of course, this will almost certainly also be the year of the “AI smartphone” — that is to say handsets leveraging platforms’ GPT models from companies like OpenAI, Google and Microsoft to bring new methods for interacting with consumer devices. Meanwhile, upstart rabbit generated buzz last month at CES for its own unique take on the generative AI-first consumer device.

For its part, Humane has a lot riding on this launch. The company has thus far raised around $230 million, including last year’s $100 million Series C. There’s a lot to be said for delaying a product until it’s consumer ready. While early adopters are — to an extent — familiar with first-gen bugs, there’s always a limit to such patience. At the very least, a product like this will need to do most of what it’s supposed to do most of the time.

During CES, the company announced that it had laid off 10 employees, amounting to 10% of its total workforce. That’s not a huge number for a startup of that size, but it’s absolutely notable when it occurs at a well-funded company at a time when it needs to project confidence to consumers and investors, alike.

The Ai Pin is currently available for preorder at $699. Those who do so prior to March 31 will get three months of the device’s $24/month subscription service for free.





Source link

Continue Reading

Trending

Copyright © 2023 Dailycrunch. & Managed by Shade Marketing & PR Agency