Technology

OpenAI found features in AI models that correspond to different ‘personas’

OpenAI researchers say they’ve discovered hidden features inside AI models that correspond to misaligned “personas,” or types of people, according to new research published by the company on Wednesday.

By looking at an AI model’s internal representations — the numbers that dictate how an AI model responds, which often seem completely incoherent to humans — OpenAI researchers were able to find patterns that lit up when a model misbehaved.

The researchers found one such feature that corresponded to toxic behavior in an AI model’s responses — meaning the AI model would lie to users or make irresponsible suggestions, like asking the user to share their password or hack into a friend’s account.

The researchers discovered they were able to turn toxicity up or down simply by adjusting the feature.

OpenAI’s latest research gives the company a better understanding of the factors that can make AI models act unsafely, and thus, could help them develop safer AI models. OpenAI could potentially use the patterns they’ve found to better detect misalignment in production AI models, according to OpenAI interpretability researcher Dan Mossing.

“We are hopeful that the tools we’ve learned — like this ability to reduce a complicated phenomenon to a simple mathematical operation — will help us understand model generalization in other places as well,” said Mossing in an interview with TechCrunch.

AI researchers know how to improve AI models, but confusingly, they don’t fully understand how AI models arrive at their answers — Anthropic’s Chris Olah often remarks that AI models are grown more than they are built. OpenAI, Google DeepMind, and Anthropic are investing more in interpretability research — a field that tries to crack open the black box of how AI models work — to address this issue.

A recent study from independent researcher Owain Evans raised new questions about how AI models generalize. The research found that OpenAI’s models could be fine-tuned on insecure code and would then display malicious behaviors across a variety of domains, such as trying to trick a user into sharing their password. The phenomenon is known as emergent misalignment, and Evans’ study inspired OpenAI to explore this further.

But in the process of studying emergent misalignment, OpenAI says it stumbled into features inside AI models that seem to play a large role in controlling behavior. Mossing says these patterns are reminiscent of internal brain activity in humans, in which certain neurons correlate to moods or behaviors.

“When Dan and team first presented this in a research meeting, I was like, ‘Wow, you guys found it,’” said Tejal Patwardhan, an OpenAI frontier evaluations researcher, in an interview with TechCrunch. “You found like, an internal neural activation that shows these personas and that you can actually steer to make the model more aligned.”

Some features OpenAI found correlate to sarcasm in AI model responses, whereas other features correlate to more toxic responses in which an AI model acts as a cartoonish, evil villain. OpenAI’s researchers say these features can change drastically during the fine-tuning process.

Notably, OpenAI researchers said that when emergent misalignment occurred, it was possible to steer the model back toward good behavior by fine-tuning the model on just a few hundred examples of secure code.

OpenAI’s latest research builds on the previous work Anthropic has done on interpretability and alignment. In 2024, Anthropic released research that tried to map the inner workings of AI models, trying to pin down and label various features that were responsible for different concepts.

Companies like OpenAI and Anthropic are making the case that there’s real value in understanding how AI models work, and not just making them better. However, there’s a long way to go to fully understand modern AI models.

Technology

Pintarnya raises $16.7M to power jobs and financial services in Indonesia

Pintarnya, an Indonesian employment platform that goes beyond job matching by offering financial services along with full-time and side-gig opportunities, said it has raised a $16.7 million Series A round.

The funding was led by Square Peg with participation from existing investors Vertex Venture Southeast Asia & India and East Ventures.

Ghirish Pokardas, Nelly Nurmalasari, and Henry Hendrawan founded Pintarnya in 2022 to tackle two of the biggest challenges Indonesians face daily: earning enough and borrowing responsibly.

“Traditionally, mass workers in Indonesia find jobs offline through job fairs or word of mouth, with employers buried in paper applications and candidates rarely hearing back. For borrowing, their options are often limited to family/friend or predatory lenders with harsh collection practices,” Henry Hendrawan, co-founder of Pintarnya, told TechCrunch. “We digitize job matching with AI to make hiring faster and we provide workers with safer, healthier lending options — designed around what they can reasonably afford, rather than pushing them deeper into debt.”

Around 59% of Indonesia’s 150 million workforce is employed in the informal sector, highlighting the difficulties these workers encounter in accessing formal financial services because they lack verifiable income and official employment documentation.

Pintarnya tackles this challenge by partnering with asset-backed lenders to offer secured loans, using collateral such as gold, electronics, or vehicles, Hendrawan added.

Since its seed funding in 2022, the platform currently serves over 10 million job seeker users and 40,000 employers nationwide. Its revenue has increased almost fivefold year-over-year and expects to reach break-even by the end of the year, Hendrawn noted. Pintarnya primarily serves users aged 21 to 40, most of whom have a high school education or a diploma below university level. The startup aims to focus on this underserved segment, given the large population of blue-collar and informal workers in Indonesia.

Techcrunch event

San Francisco
|
October 27-29, 2025

“Through the journey of building employment services, we discovered that our users needed more than just jobs — they needed access to financial services that traditional banks couldn’t provide,” said Hendrawan. “We digitize job matching with AI to make hiring faster and we provide workers with safer, healthier lending options — designed around what they can reasonably afford, rather than pushing them deeper into debt.”

Pintarnya Founders Press Kit — image credits: Pintarnya (Founders of Pintarnya Ghirish Pokardas, Nelly Nurmalasari, and Henry Hendrawan)

While Indonesia already has job platforms like JobStreet, Kalibrr, and Glints, these primarily cater to white-collar roles, which represent only a small portion of the workforce, according to Hendrawan. Pintarnya’s platform is designed specifically for blue-collar workers, offering tailored experiences such as quick-apply options for walk-in interviews, affordable e-learning on relevant skills, in-app opportunities for supplemental income, and seamless connections to financial services like loans.

The same trend is evident in Indonesia’s fintech sector, which similarly caters to white-collar or upper-middle-class consumers. Conventional credit scoring models for loans, which rely on steady monthly income and bank account activity, often leave blue-collar workers overlooked by existing fintech providers, Hendrawan explained.

When asked about which fintech services are most in demand, Hendrawan mentioned, “Given their employment status, lending is the most in-demand financial service for Pintarnya’s users today. We are planning to ‘graduate’ them to micro-savings and investments down the road through innovative products with our partners.”

The new funding will enable Pintarnya to strengthen its platform technology and broaden its financial service offerings through strategic partnerships. With most Indonesian workers employed in blue-collar and informal sectors, the co-founders see substantial growth opportunities in the local market. Leveraging their extensive experience in managing businesses across Southeast Asia, they are also open to exploring regional expansion when the timing is right.

“Our vision is for Pintarnya to be the everyday companion that empowers Indonesians to not only make ends meet today, but also plan, grow, and upgrade their lives tomorrow … In five years, we see Pintarnya as the go-to super app for Indonesia’s workers, not just for earning income, but as a trusted partner throughout their life journey,” Hendrawan said. “We want to be the first stop when someone is looking for work, a place that helps them upgrade their skills, and a reliable guide as they make financial decisions.”

Technology

OpenAI warns against SPVs and other ‘unauthorized’ investments

In a new blog post, OpenAI warns against “unauthorized opportunities to gain exposure to OpenAI through a variety of means,” including special purpose vehicles, known as SPVs.

“We urge you to be careful if you are contacted by a firm that purports to have access to OpenAI, including through the sale of an SPV interest with exposure to OpenAI equity,” the company writes. The blog post acknowledges that “not every offer of OpenAI equity […] is problematic” but says firms may be “attempting to circumvent our transfer restrictions.”

“If so, the sale will not be recognized and carry no economic value to you,” OpenAI says.

Investors have increasingly used SPVs (which pool money for one-off investments) as a way to buy into hot AI startups, prompting other VCs to criticize them as a vehicle for “tourist chumps.”

Business Insider reports that OpenAI isn’t the only major AI company looking to crack down on SPVs, with Anthropic reportedly telling Menlo Ventures it must use its own capital, not an SPV, to invest in an upcoming round.

Technology

Meta partners with Midjourney on AI image and video models

Meta is partnering with Midjourney to license the startup’s AI image and video generation technology, Meta Chief AI Officer Alexandr Wang announced Friday in a post on Threads. Wang says Meta’s research teams will collaborate with Midjourney to bring its technology into future AI models and products.

“To ensure Meta is able to deliver the best possible products for people it will require taking an all-of-the-above approach,” Wang said. “This means world-class talent, ambitious compute roadmap, and working with the best players across the industry.”

The Midjourney partnership could help Meta develop products that compete with industry-leading AI image and video models, such as OpenAI’s Sora, Black Forest Lab’s Flux, and Google’s Veo. Last year, Meta rolled out its own AI image generation tool, Imagine, into several of its products, including Facebook, Instagram, and Messenger. Meta also has an AI video generation tool, Movie Gen, that allows users to create videos from prompts.

The licensing agreement with Midjourney marks Meta’s latest deal to get ahead in the AI race. Earlier this year, CEO Mark Zuckerberg went on a hiring spree for AI talent, offering some researchers compensation packages worth upwards of $100 million. The social media giant also invested $14 billion in Scale AI, and acquired the AI voice startup Play AI.

Meta has held talks with several other leading AI labs about other acquisitions, and Zuckerberg even spoke with Elon Musk about joining his $97 billion takeover bid of OpenAI (Meta ultimately did not join the offer, and OpenAI denied Musk’s bid).

While the terms of Meta’s deal with Midjourney remain unknown, the startup’s CEO, David Holz, said in a post on X that his company remains independent with no investors; Midjourney is one of the few leading AI model developers that has never taken on outside funding. At one point, Meta talked with Midjourney about acquiring the startup, according to Upstarts Media.

Midjourney was founded in 2022 and quickly became a leader in the AI image generation space for its realistic, unique style. By 2023, the startup was reportedly on pace to generate $200 million in revenue. The startup sells subscriptions starting at $10 per month. It offers pricier tiers, which offer more AI image generations, that cost as much as $120 per month. In June, the startup released its first AI video model, V1.

Techcrunch event

San Francisco
|
October 27-29, 2025

Meta’s partnership with Midjourney comes just two months after the startup was sued by Disney and Universal, alleging that it trained AI image models on copyrighted works. Several AI model developers — including Meta — face similar allegations from copyright holders, however, recent court cases pertaining to AI training data have sided with tech companies.

Got a sensitive tip or confidential documents? We’re reporting on the inner workings of the AI industry — from the companies shaping its future to the people impacted by their decisions. Reach out to Rebecca Bellan at [email protected] and Maxwell Zeff at [email protected]. For secure communication, you can contact us via Signal at @rebeccabellan.491 and @mzeff.88.

We’re always looking to evolve, and by providing some insight into your perspective and feedback into TechCrunch and our coverage and events, you can help us! Fill out this survey to let us know how we’re doing and get the chance to win a prize in return!