Techin' Geek

Mercor, an AI recruiting firm of notable reputation, has reported a security breach linked to a supply chain attack involving the open-source initiative LiteLLM.

On Tuesday, the AI company informed TechCrunch that it was “one of numerous firms” impacted by a recent breach of the LiteLLM project, which has ties to a hacking collective known as TeamPCP. The confirmation of this breach comes as the extortion hacker group Lapsus$ claimed to have targeted Mercor and accessed its information.

It is currently uncertain how the Lapsus$ group acquired the compromised data from Mercor amid TeamPCP’s cyber assault.

Established in 2023, Mercor collaborates with organizations such as OpenAI and Anthropic to enhance AI models by hiring specialized domain experts, including scientists, doctors, and lawyers from regions such as India. The company claims to facilitate over $2 million in daily transactions and was assessed at $10 billion following a $350 million Series C funding round led by Felicis Ventures in October 2025.

Mercor representative Heidi Hagberg confirmed to TechCrunch that the firm had “acted swiftly” to address and resolve the security issue.

“We are undergoing a comprehensive investigation aided by top third-party forensic specialists,” Hagberg stated. “We will maintain open communication with our clients and contractors directly as deemed appropriate and allocate the necessary resources to address the situation as quickly as possible.”

Previously, Lapsus$ took responsibility for the suspected data breach on its leaking platform and provided a sample of data supposedly obtained from Mercor, which TechCrunch examined. The sample contained material related to Slack data and what seemed to be ticketing information, along with two videos allegedly depicting dialogues between Mercor’s AI frameworks and contractors on its platform.

Hagberg chose not to respond to follow-up inquiries regarding whether the incident was associated with the assertions made by Lapsus$, or whether any data belonging to customers or contractors had been accessed, exfiltrated, or misappropriated.

The compromise of LiteLLM initially came to light last week after malicious code was identified in a package tied to the Y Combinator-supported startup’s open-source project. While the harmful code was detected and removed within hours, the event raised concerns due to LiteLLM’s extensive adoption online, with the library downloaded millions of times daily, according to security firm Snyk. The occurrence also led LiteLLM to implement adjustments in its compliance measures, including a switch from the controversial startup Delve to Vanta for compliance certifications.

It is still unclear how many businesses were impacted by the LiteLLM-related incident or if any data exposure took place, as investigations are ongoing.

At times, Michiko Kato reflects on the significant obstacles she has faced in her life. 

She once left everything behind in her hometown of Tokyo to enroll at Harvard Business School in Massachusetts. One day, she chose to step away from her secure career in finance to plunge into the uncertain world of startups. That leap is what altered her path entirely. 

On Wednesday, Kato officially takes on her most significant challenge to date — CIO of Toyota’s Woven Capital and CEO of Toyota Invention Partners. This role makes her the first female CEO of a wholly owned subsidiary of Toyota. 

Woven Capital serves as the growth-stage venture capital division of Toyota, dedicated to supporting founders innovating in mobility (which includes areas like space, cybersecurity, and autonomous vehicles). The company recently announced an $800 million Fund II last August (Fund I, also at $800 million, began in 2021), with intentions to finance at least 20 new Series B investments. Its portfolio includes companies like satellite provider Xona and the defense manufacturing firm Machina Labs. 

The firm aims to discover the “future leader of mobility,” she mentioned, and seeks to select companies that will serve as “collaboration partners with Toyota.”

“We can co-lead, make smaller investments, or undertake aggressive investing; we strive to remain adaptable,” she stated. And regarding her approach? “I aim to be involved. I want to bring value to startups. And, I want to concentrate on building partnerships.” 

This week, Kato is not the only one receiving a promotion at the firm. Mia Panzer is transitioning from her role in business strategy at one of Toyota’s technology subsidiaries to become COO of Woven Capital. This means that two key positions at this corporate VC (CVC) firm will be occupied by women, marking a notable shift in a traditionally male-dominated finance and investment industry.

Historically, women have experienced (a slight advantage) in ascending the ranks at CVCs compared to traditional venture firms. However, the latest concrete data stems from a 2014 CBI report indicating that just under 20% of leading CVCs included women on their investment teams. This figure was contrasted at the time with the mere 7% of partners at the top 100 venture firms being women.

Currently, in venture capital, that percentage is around 15.4%, suggesting that the proportion of women in investment roles at CVCs has also risen. 

Kato joined the company back in 2020 as one of the first recruits for what was then the newly established Woven Capital. (It was spun out from an internal subsidiary of Toyota.) 

She has amassed 15 years of investing experience overall, including time spent on the M&A team at Unison Capital and serving as CFO for the Japanese AI startup ABEJA. Since coming on board at Woven, she has spearheaded six investments (including one undisclosed) in startups such as the reusable rocket venture Stoke and the autonomous vehicle company Nuro, marking her initial investment. She is particularly passionate about aeromobility, physical AI, and hardware. “I believe we can fundamentally redefine the manufacturing process,” she shared regarding her vision for the CVC. 

Collaborating with her will be Panzer, in the newly created COO position. She will oversee finance, operations, HR, and legal strategy. She noted that there are two main concerns every CVC has — a corporate slowdown that hinders deals and misalignment with the parent company. “My role is essentially to help navigate this,” she stated. 

She has been partnering with Ro Gupta, Woven Capital’s managing director, since his days of establishing the mapping firm Camera in 2019. She joined that startup while three months pregnant with her first child to lead finance. Before that, she was at Goldman and then at the pet wellness company Independent Pet Partners (IPP). She explained that she was not actively seeking to leave her position at IPP but decided to give Gupta and Camera a try. 

“I genuinely enjoy being engaged in the startup ecosystem,” she expressed. 

Toyota’s tech division subsequently acquired Carmera (a sale she played a role in facilitating), resulting in Gupta and Panzer joining that sector. In December 2025, he took on the role of managing director of Woven Capital and chose to bring Panzer along with him. Kato and Panzer report to Gupta.

Initially, Panzer didn’t believe she was fully equipped for the role, but then remembered how often women feel they need to meet every qualification for a job, while men tend to dive right in. She recalled a career filled with assumptions about her capabilities, from a college recruiter doubting her technical expertise because she was female to a Goldman colleague saying she didn’t see her as competition for promotion due to her gender. 

She resolved to take the plunge as well.

“We discuss a lot about the Japanese concept of ikigai, where one seeks to identify what they are good at, what they love, what the world requires, and what they can earn a living from,” she stated, noting that it feels like a full circle moment as generations of her family have worked in automotive parts. “I always advise other women, ‘Let them have low expectations,” she said. “It’s easy to exceed them.’” 

Anthropic has established its public persona around the notion of being the diligent AI firm. It shares comprehensive research on AI risks, employs top-tier researchers in the domain, and has been outspoken regarding the obligations tied to creating such potent technology — so outspoken, in fact, that it is currently engaged in a dispute with the Department of Defense. On Tuesday, regrettably, someone overlooked checking a box.

This marks the second occasion within a week. Last Thursday, Fortune revealed that Anthropic had unintentionally made almost 3,000 internal documents accessible to the public, comprising a draft blog entry detailing a significant new model the company had yet to unveil.

Here’s what transpired on Tuesday: When Anthropic deployed version 2.1.88 of its Claude Code software suite, it inadvertently incorporated a file that revealed nearly 2,000 source code files and exceeding 512,000 lines of code — effectively the complete architectural design for one of its key products. A security researcher named Chaofan Shou recognized the issue almost instantaneously and shared it on X. Anthropic’s response to various media outlets was somewhat laid-back as far as such incidents go: “This was a release packaging issue triggered by human error, not a security compromise.” (Internally, one might speculate that the tone was less measured.)

Claude Code is far from a trivial product. It’s a command-line tool that enables developers to use Anthropic’s AI for coding and editing tasks and has grown robust enough to disturb competitors. According to the WSJ, OpenAI discontinued its video generation service Sora merely six months post-launch to redirect its focus towards developers and enterprises — in part due to Claude Code’s rising influence.

What was exposed was not the AI model per se but the software framework surrounding it — the directives that inform the model’s behavior, the tools it should utilize, and its boundaries. Developers began releasing in-depth evaluations almost immediately, with one characterizing the product as “a production-grade developer experience, not merely an interface for an API.”

Whether this will have any significant long-term implications is a question best posed to developers. Competitors may glean insights from the architecture; meanwhile, the industry is evolving rapidly.

Regardless, somewhere at Anthropic, one can envision a highly skilled engineer spending the remainder of the day quietly questioning their job security. One can only wish it’s not the same engineer, or engineering team, from the previous week.

Salesforce, the cloud computing leader, has been transforming its business with AI, and at a small event in San Francisco on Tuesday, CEO Marc Benioff and his team revealed the latest outcomes of those initiatives: a refreshed version of Slack, packed with a multitude of new AI functionalities. The standout feature is a major enhancement for its AI assistant, Slackbot.

The 30 new features, set to be rolled out in the upcoming months, follow a January upgrade that granted Slackbot agent-like abilities — such as drafting emails, scheduling meetings, and filtering through your inbox for specific details.

Perhaps the highlight feature introduced on Tuesday is what the company refers to as reusable AI-skills — enabling users to establish specific tasks for Slackbot that, once configured, can be utilized across various situations and contexts. Salesforce indicates that Slackbot includes a built-in library of AI-skills, while users can also create their own customized versions.

After these skills are established, they considerably lessen the workload an employee may face. For instance, a user can activate a skill through a simple command in Slack — like “create a budget” for an upcoming event — prompting Slackbot to gather all pertinent information from the company’s Slack channels and any connected apps or data sources, producing a workable plan. The bot will then autonomously arrange a meeting to discuss the plan, inviting relevant staff based on their positions.

Slackbot now additionally acts as an MCP (Model Context Protocol) client — signifying it can integrate with and collaborate with external services and tools. Among these is Agentforce, Salesforce’s AI agent development platform introduced in 2024. Through this connection, it can “route work or prompt queries to Agentforce or any agent or app in your enterprise,” according to the company, with the agent determining the most relevant and efficient pathway for the information, without any human involvement.

As stated by Rob Seaman, Slack’s interim CEO and former chief product officer, Slackbot is now also capable of transcribing meetings and summarizing them. If a participant in a meeting drifts off, thereby missing essential details, they can simply request Slackbot to provide a summary of the meeting, including any action items assigned to them.

The assistant can also now function outside of Slack and keep track of your desktop activities — Salesforce cites “your deals, your conversations, your calendar, and your habits” as data sources it utilizes. Based on that context, the bot will offer actionable suggestions or draft follow-ups for important tasks. Seaman has noted that privacy safeguards are incorporated into this design, allowing users to modify permissions as required.

In summary: Salesforce is evidently striving to elevate Slack beyond its original function as an enterprise communication tool and position it as a more adaptable platform capable of managing a broader range of business tasks. The aim appears to be that by integrating AI extensively, Slack can evolve into an essential component of enterprise users’ key business operations.

Benioff allowed his team to present the main features on Tuesday but noted in his keynote that the five years since Salesforce acquired Slack have been an “incredible journey,” one that has resulted in “two and a half times revenue growth.” He added: “We have approximately a million businesses operating on Slack. It’s been a significant growth narrative.”

In February, Senator Ed Markey (D-MA) sent inquiries to seven U.S. firms engaged in autonomous vehicle technology along with a set of questions. He particularly sought to understand how frequently these firms’ vehicles — operated by Aurora, May Mobility, Motional, Nuro, Tesla, Waymo, and Zoox — depend on feedback from remote personnel. According to the outcomes of Markey’s investigation released on Tuesday, they all declined to provide answers.

The data published by Markey’s office exemplifies the reluctance of autonomous vehicle companies to disclose specifics regarding their operational methods — even as they test this technology on public roadways.

“This report has uncovered a remarkable absence of transparency from the AV companies concerning their use of [remote assistance operators] to assist in directing their AVs. The investigation revealed a patchwork of safety protocols across the sector, with notable discrepancies in operator qualifications, response times, and international staffing, all in the absence of any federal guidelines regulating these operations,” Markey’s office stated in its report.

Markey announced on Tuesday that he is urging the National Highway Traffic Safety Administration to examine these firms’ employment of remote assistance personnel, and that he is “developing legislation to enforce stringent regulations on AV companies’ utilization of remote operators.”

TechCrunch has contacted each company mentioned. Waymo opted not to comment. The other six did not respond immediately.

Markey initiated his investigation in February following a Senate Commerce Committee hearing focused on the future of self-driving vehicles. At that hearing, Waymo’s chief safety officer Mauricio Peña discussed how the company’s vehicles occasionally require direction from “remote assistance” staff when encountering challenging or unanticipated situations. Peña also disclosed that roughly half of Waymo’s remote assistance team is situated in the Philippines.

Over the years, autonomous vehicle companies have intermittently addressed these types of remote assistance operations. However, those discussions were often speculative, as the technology remained in its early testing stages.

Now that numerous companies have commercially introduced robotaxis or, in the case of Aurora, self-driving trucks, scrutiny of their complete operations has increased.

After the hearing, Markey dispatched letters to those seven firms seeking additional information about their remote operations. His office posed 14 questions to each company, including the frequency with which remote staff provide guidance to autonomous vehicles, the size of these teams, their locations, licensing, and the security protocols they implement.

The companies’ responses — which can be read in full here — are highly varied. None directly addressed the query about how often their remote staff assist AVs, with Waymo and May Mobility specifically stating that this constitutes “confidential business information.” Tesla did not even include this question in its reply letter. The reason remains unclear, as the company has not had a North American communications team for years.

Waymo asserted in its response that enhancements to its self-driving system have “materially reduced” the number of help requests per mile that its vehicles transmit to remote staff, but it did not provide any specifics or evidence. The company mentioned that a “vast majority of requests” sent by its robotaxis to remote assistance staff are resolved by the self-driving system “before an agent even provides a response.”

Waymo was the sole company to acknowledge the use of international remote assistance personnel. While the company asserts that it ensures these workers have local drivers’ licenses, Markey’s office noted on Tuesday that a “driver’s license in a foreign location is not a substitute for passing a U.S. driver’s license examination, as road rules will almost certainly differ by region.”

All companies except Tesla contended that they either prohibit or lack the capability for remote assistance personnel to directly operate these autonomous vehicles. In contrast, Tesla claimed that its remote assistance workers “are authorized to temporarily assume direct vehicle control as the final course of action after exhausting all other available interventions.”

Tesla stated that this can only occur if a vehicle in its pilot fleet is moving at 2 miles per hour or less, and that the remote operator is restricted to controlling the vehicle at a maximum speed of 10 miles per hour.

“This capability allows Tesla to swiftly relocate a vehicle that might be in a precarious position, thus reducing the need to wait for a first responder or Tesla field representative to manually retrieve the vehicle,” the company communicated to Markey’s office.

This practice has recently attracted criticism for Waymo, which faced challenging inquiries from San Francisco city officials at a hearing this month regarding its dependency on first responders for moving incapacitated robotaxis. Waymo does possess its dedicated “roadside assistance” team, distinctly separate from its remote assistance personnel, as detailed by TechCrunch recently. However, this aspect of Waymo’s operation was not a primary focus of Markey’s investigation.

Markey’s office did extract some additional information from these companies. His report details the latency observed in these remote assistance interactions (which fluctuates by company, with May Mobility citing the longest worst-case scenario at 500 milliseconds), how certain companies attempt to keep these workers from becoming fatigued, and what measures they adopt to safeguard the data they manage.

These are inquiries that autonomous vehicle companies have dealt with for years, and obtaining answers has not been straightforward. However, with numerous commercial deployments imminent, Markey’s office will certainly not be the last to demand more information.

OpenAI has finalized an agreement to secure $122 billion at a valuation of $852 billion, marking its most significant funding round to date as the firm prepares to enter public markets this year.

This round will enhance OpenAI’s financial resources as it allocates substantial funds towards AI chips, construction of data centers, and recruitment of top-tier talent.

SoftBank co-led the round with Andreessen Horowitz, D.E. Shaw Ventures, MGX, TPG, and T. Rowe Price Associates, with contributions from Amazon, Nvidia, and Microsoft.

Approximately $3 billion was sourced from individual investors through bank channels. OpenAI will also feature in various ETFs managed by ARK Invest, providing wider access to its stock and expanding its shareholder base ahead of its anticipated IPO.

OpenAI mentioned that it has increased its revolving credit facility to about $4.7 billion, backed by several leading global banks. The company indicated that the facility remains undrawn, suggesting it is enhancing its financial flexibility as it increases spending on compute and infrastructure, rather than addressing immediate liquidity requirements.

The press release regarding the funding reads more like a preliminary S-1 draft than a traditional blog post; it is rich in flywheel metaphors, delves into revenue per compute unit, and includes the kind of TAM-justifying rhetoric that captivates institutional investors.

OpenAI provided updates on its revenue and user metrics, asserting it is generating $2 billion monthly and taking aim at rivals: “Currently, we are increasing revenue at four times the rate of the companies that pioneered the Internet and mobile eras, such as Alphabet and Meta.”

The firm also reported having over 900 million weekly active users in consumer AI and more than 50 million subscribers, with search usage nearly tripling within the past year. OpenAI noted that its ads pilot is yielding over $100 million in annual recurring revenue in less than six weeks, presenting a significant potential revenue avenue for a company that established its user base without advertising.

The AI powerhouse claims that business momentum is reflected in its operations, now accounting for 40% of its revenue (up from roughly 30% last year) and is “on track to match consumer by the end of 2026.” The company attributed its growth in agentic workflows to its latest model, GPT-5.4.

In conclusion, OpenAI referred to itself as an “AI superapp,” clearly indicating its ambition to dominate the main interface through which people interact with AI.

Overall, this conveys a single message: OpenAI is actively crafting its public market story, and this funding round is as much about solidifying IPO expectations as it is about securing capital.

Pete Martin recalls securing a $5 million seed round at a $25 million post-money valuation for his AI-driven cybersecurity firm Realm back in 2024, which feels like a millennium in “AI years”.  

At that time, that valuation appeared elevated for that sum, he noted. However, presently, “it’s quite standard” to observe a $10 million seed round at a $40 million to $45 million post-money valuation, particularly for an AI enterprise, he mentioned.  

In fact, this tends to occur exclusively for AI firms, as investors are showing scant interest in any other sectors. 

During the latest Y Combinator Demo Day in March, conversations revolved around the inflated valuations of companies, stated Ashley Smith, a general partner at early-stage fund Vermilion. Numerous startups had clinched six- to seven-figure customer contracts, including one that had only been operational for eight weeks, she remarked, leading to companies requesting $5 million at a $40 million post-money valuation.

This situation surpassed the so-called “YC tax,” which illustrates how much extra investors are prepared to invest simply because the startup is a YC alumnus, she explained. Even with those early revenue figures, Smith asserted that market investors are valuing rounds “years ahead of actual traction.”

Major venture firms, buoyed with capital, are also entering rounds at earlier stages, escalating startup prices and valuations with the hope of attaining significant returns if these companies eventually go public or exit. Smaller VC firms also exhibit an unquenchable thirst for AI enterprises. As an investor concentrated on AI infrastructure, Smith noted that she frequently finds herself priced out of a round, especially when a larger firm steps in. This is one reason both founders and VCs claim seed deal volume has declined, even as valuations have soared, according to data from Carta.  

Shanea Leven, founder of the enterprise AI application platform Empromptu, points the finger at Cursor, which achieved $100 million in revenue within just 12 months in early 2025. It was one of the first prominent AI companies to raise the standard for how rapidly these startups could achieve traction, although it certainly wasn’t the sole example. Others include Lovable, Bolt, OpenEvidence, and ElevenLabs, all touting their rapid progress. Although these are exceptions, it’s challenging for some not to feel the generated excitement.  

“The expectations from investors are changing now,” she claimed. “The pressure is unprecedented, not just to become a billion-dollar entity, but a $50 billion one.”   

Accelerated traction, larger valuations 

VCs are quick to justify the rise in seed valuations. For example, Marlon Nichols, managing general partner at MaC Ventures, stated that the evidence lies in the traction evident right from the start, influencing seed pricing. When he founded his firm in 2019, he noted his average entry investment was $2.5 million. Currently, it’s $5 million.  

“The top seed-stage firms no longer resemble typical seed-stage companies,” he asserted. The evolution of AI tools enables founders to reach minimal viable products and acquire early customers more rapidly than ever, even among large enterprises that are eagerly scouting for AI deployment solutions.

Nichols’ last two seed investments were generating over $2 million in revenue, featuring “paid pilots from substantial enterprises” and “a clear pathway to full commercial agreements.” He issued checks ranging between $3 million and $4 million and agreed to value the startups at $25 million and $30 million post-money, respectively, significantly more compared to a few years ago.  

The founders’ past experiences also impacted his term-sheet offers. “They possessed relevant backgrounds” and “a history of execution,” he explained, “which mitigated much of that early-stage risk.”  

Moreover, investors are prepared to pay premium prices for proven AI talent, favoring second-time founders or those possessing credentials from recognized prior employers (such as OpenAI). This also raises anticipated valuations across the landscape.

“There’s an ongoing competition for outstanding researchers currently, and I don’t classify it as good or bad; it simply reflects the existing market conditions,” Amber Atherton, a partner at the early-stage consumer fund Patron, noted.  

That’s what is driving some of the most extreme seed valuations, like the $2 billion seed for Thinking Machine Labs at a $12 billion valuation by ex-OpenAI Mira Murati.  

Leven, who is a second-time founder, mentioned that her startup’s valuation at this stage is double that of her initial company at a comparable phase. Not only is her latest venture AI-focused, but it also has substantially more traction than her previous startup did at this point, illustrating how swiftly new businesses like hers can expand.  

“I presently hold several six-figure contracts and am about to close a seven-figure deal. You need these to secure funding,” Leven explained. “A friend of mine is attempting to raise a comparable amount, but hers is not in AI, and it took her two years to achieve half of what I secured in three weeks.”  

Pre-seed is the new seed 

Seed VCs like Vermilion’s Smith are countering increasing seed valuations by pursuing more pre-seed deals. Pre-seed startups resemble the kind of companies that seed firms used to be many years ago: extremely early and pre-revenue. 

Jonathan Lehr, a general partner at Work-Bench, is investing from a $160 million fund predominantly aimed at seed rounds, though he mentioned that the firm has become “more comfortable” participating at the pre-seed stage as companies scale much quicker.  

It’s becoming more typical to see investors inject capital into startups at earlier stages, as enhanced exposure is simply the cost of “accessing firms that can scale rapidly and emerge as category leaders,” Lehr commented. 

Meanwhile, Atherton noted that to secure a stake in these promising early-stage enterprises, the average check size for her firm’s $100 million Fund II now spans from $4 million to $5 million, up from $1 to $2 million for its $90 million Fund I.

“AI has significantly elevated the standards for founders to launch with live products and customers right away,” she stated.  “Investors must act more swiftly and assess real-world traction at an earlier stage since the best founders are delivering products with users and revenue almost immediately.”

Thus, seed VCs are no longer “backing ideas”; they are “backing early indications of genuine consumer product demand,” she articulated. Seed VCs are also accelerating their pace, shifting “from slow diligence to high-conviction judgments regarding distribution, retention, and founder characteristics.” 

But there’s a catch

As expectations have risen, so have investors’ demands.  

Atherton expressed that it’s no longer adequate for a company to merely develop and dispatch a product. Today, anyone can achieve that. It’s not solely about traction, although that is beneficial. It centers on the future, the narrative founders present about how they will outperform their competitors and dominate the market. This is what these seed VCs believe will propel startups towards sustained, $50 billion+ valuations or at least lead to some profitable exit.

“Individuals are merely striving to withstand the pressure,” Leven noted. “Otherwise, there won’t be sufficient funds to expand and genuinely compete.”  

The upside of raising significant amounts of capital at the nascent stages for a founder is that it enables the company to accelerate and recruit costly talent. VCs understand, as they formulate their term sheets, that talent during the AI era is expensive, as are the operations of the AI models supporting these startups, and contending with other well-funded adversaries, sometimes large SaaS companies already valued in the billions. 

Everyone, according to Leven, is attempting to replicate the success of Google’s acquisition of Wiz. However, the stakes are higher as well. Founders must evolve their businesses into entities that validate the elevated early valuations before seeking further funding. Series A investors are also anticipating larger, quicker, and more.  

Nichols and his firm are currently evaluating an increasing number of young companies, with new expectations that they meet their milestones within approximately 18 months. “That discipline is equally crucial as backing winners,” he stated.  

Elevated seed valuations lead to a narrower margin for mistakes, Lehr commented, adding: “Less latitude for experimentation, diminished tolerance for pivots, and heightened scrutiny if progress doesn’t align with the capital raised.”  

Martin, the cybersecurity entrepreneur, successfully completed his Series A funding late last year, mentioning that the benchmarks were manageable for his firm to meet. Yet, he too cautioned founders.

“You might find yourself trapped in the middle,” Martin warned. “Too costly for new investors, yet lacking the momentum to validate the next round.”

At times, what seems like a solid concept, a substantial investment from a prominent VC, and a host of well-connected angel investors may not suffice.

Fewer than twelve months post-launch, Yupp is shutting down, co-founders Pankaj Gupta and Gilad Mishne revealed on Tuesday.

Yupp provided a crowdsourced service for selecting AI models. It permitted users to experiment with and evaluate outcomes from a collection of 800 AI models for free, including cutting-edge options from OpenAI, Google, and Anthropic. Yupp would deliver various responses to users’ prompts, providing information or images, with users giving feedback on which models were most effective and the reasons behind their choices.

The objective was to compile anonymized insights on actual consumer needs regarding AI that model developers would then purchase. Yupp claimed to have acquired 1.3 million users and gathered millions of preferences monthly. It even maintained a leaderboard. The company noted it had a few AI labs among its clients.

Unfortunately, it “didn’t achieve a sufficiently robust product-market fit” to remain viable, partly due to rapid advancements in AI models over recent months, as stated by the founders.

While labs are investing significantly for insights, the prevailing model — established by firms like Scale AI and Mercor — involves employing specialized experts, such as PhDs, to integrate into the reinforcement learning paradigm.

Moreover, Silicon Valley is already strategizing for the future, envisioning a time when AI systems are designed for, and utilized by, other AIs. Although model developers may seek consumer insights at present, they are primarily focused on the era when agents, rather than humans, dominate the digital landscape.

“The landscape of AI model capabilities has transformed significantly in just the past year and will keep evolving rapidly,” Gupta, Yupp’s CEO, stated in a post on X regarding the shutdown plans. “The future encompasses not merely models but agentic systems.”

Yupp secured a $33 million seed funding round in 2024, spearheaded by a16z crypto’s Chris Dixon, a substantial seed round for its time. Additionally, Yupp received investments from over 45 angels and small investors, it mentioned. This included notable figures such as Google DeepMind chief scientist Jeff Dean; Twitter co-founder Biz Stone; Pinterest co-founder Evan Sharp; and Perplexity CEO Aravind Srinivas.

Gupta indicated that some of Yupp’s staff are joining a “well-known” AI company, while others are seeking new opportunities. Yupp did not promptly respond to TechCrunch’s inquiry for a statement.

Meta faced a setback in a lawsuit against the state of New Mexico last week, signifying the first occasion the enterprise has been deemed accountable by the judicial system for jeopardizing child safety. This ruling was significant in its own right — but the following day, Meta encountered another defeat when a jury in Los Angeles concluded that the corporation deliberately crafted its applications to be addictive for minors and adolescents, thereby compromising the mental health of the plaintiff, a 20-year-old identified as K.G.M.

These decisions pave the way for an influx of lawsuits regarding Meta’s deliberate targeting of teenage users, despite its awareness that its applications can adversely affect adolescents’ mental well-being. A multitude of cases akin to K.G.M.’s are in progress, while 40 state attorneys general have initiated lawsuits against Meta that resemble New Mexico’s action.

Although social media companies are generally protected by law from liability for user-generated content, this instance focused not on the material shared on these platforms but rather on the design features themselves, such as infinite scrolling and constant notifications.

“They utilized the model that was leveraged against the tobacco industry many years ago, concentrating not on the content but on these addictive attributes — how the platform is constructed, and concerns regarding the design, which differs from content, where you encounter this First Amendment debate,” said Allison Fitzpatrick, a digital media attorney and partner at Davis+Gilbert, to TechCrunch. “It appeared to be, at least in these two instances, an effective argument.”

Following a six-week trial, the jury in the New Mexico case determined Meta guilty of breaching the state’s Unfair Practices Act, mandating the firm to pay the highest penalty of $5,000 per infringement, aggregating to a $375 million fine. The Los Angeles trial, which held Meta 70% responsible and YouTube 30% accountable for the distress experienced by plaintiff K.G.M., will impose a collective fine of $6 million on the entities involved. (Snap and TikTok reached a settlement prior to trial.)

“That amount is trivial for the Metas of the world,” Fitzpatrick remarked. “However, when you multiply that $6 million by the multitude of cases against them, it results in a staggering figure.”

“We respectfully contest these rulings and will seek to appeal,” a Meta representative informed TechCrunch. “Simplifying a complex issue like adolescent mental health to a single cause risks neglecting the many broader challenges facing youths today and disregards the reality that numerous teens depend on digital communities for connection and belonging.”

During the litigation process, newly unveiled internal documents from Meta highlighted a trend of inaction regarding the acknowledged adverse effects of its platforms on minors, as well as a focused effort to increase the amount of time teens spent on its apps, even during school hours or through “finstas,” which are “fake Instagram” accounts created by teens specifically to avoid detection by parents or teachers.

One document exhibited findings from a 2019 study, in which Meta conducted 24 in-person, individual interviews with individuals whose use of the product had been designated as problematic — a classification affecting an estimated 12.5% of users. 

“The most reliable external research suggests that Facebook’s influence on users’ well-being is negative,” the report indicates.

Several documents referenced remarks made by Meta CEO Mark Zuckerberg and Instagram head Adam Mosseri regarding their focus on engaging with teenage users. Zuckerberg even remarked that for Facebook Live to be successful among teens, his “guess is we’ll need to be very good at not notifying parents/teachers.”

In other documents, Meta staff casually discussed the company’s objectives for enhancing teen user retention.

“We discovered that one of the things we must optimize for is peeking at your phone in the middle of Chemistry :),” one employee penned in an email to Meta CPO Chris Cox.

“No one wakes up intending to maximize the number of times they check Instagram that day,” Meta VP of Product Max Eulenstein wrote in an internal email in January 2021. “Yet, that’s precisely what our product teams are striving to achieve.”

A Meta spokesperson shared with TechCrunch that many of the newly disclosed documents date back nearly a decade, but the company is attentive to feedback from parents, experts, and law enforcement on how the platform can be enhanced.

“We do not set goals related to teen time spent today,” the spokesperson stated, referencing Instagram Teen Accounts, introduced in 2024, which incorporate built-in safety measures for young users. These safeguards comprise defaulting accounts to private and permitting only those they follow to tag or mention them in posts. Instagram will also send reminders after 60 minutes of usage encouraging teens to exit the app, an adjustment that can only be made for users under 16 with parental consent. 

For Kelly Stonelake, a Director of Product Marketing at Meta, who was part of the company from 2009 to 2024, these developments are not surprising. (Stonelake is currently litigating against Meta for claimed gender-based discrimination and harassment.) 

“The substantial amount of unsealed evidence truly illustrates what I experienced firsthand,” she expressed to TechCrunch. 

At Meta, Stonelake spearheaded “go-to-market” strategies for the VR social application Horizon Worlds as it launched for teenagers. She alleges that she raised alarms over ineffective content moderation tools in the metaverse, but her concerns were dismissed.

The U.S. government has demonstrated significant interest in the topic of online safety for children, particularly following the release of damaging internal documents by Meta whistleblower Frances Haugen in 2021, revealing that Meta was aware of Instagram’s detrimental effects on adolescent girls. 

While Congress has proposed several bills aimed at enhancing children’s online safety, many privacy advocates argue that these initiatives may do more to surveil adults and restrict speech than to protect minors.

“There is no scenario where enacting censorship or ‘age verification’ legislation, under the pretense of ensuring children’s safety, won’t lead to widespread online censorship of content and speech that is deemed undesirable,” stated Fight for the Future director Evan Greer.

Stonelake previously lobbied on Capitol Hill for the Kids Online Safety Act, which has gained the most traction among these legislative efforts, receiving backing from corporations such as Microsoft, Snap, X, and Apple. However, as the bill has progressed and evolved, she has grown increasingly critical of it.

“I am advocating for a ‘no’ vote on the current iteration,” she remarked, pointing to the bill’s preemption clauses that would override state regulations concerning tech firms. “There is verbiage in the latest draft that would obstruct access to the courts for school districts, grieving families, and states — and that’s outrageous.”

This language could potentially impede the very case brought forth by New Mexico against Meta. 

“We need stakeholders to engage in discussions around solutions, rather than what’s happening now, which is merely telling a different narrative to both sides of the political aisle to incite them and instill fear,” Stonelake asserted. “The real solution will require complexity and nuance and will need to consider multiple priorities.”