16 October, 2024
Elon Musk's X could still face sanctions for training Grok on Europeans' data
Earlier this week, the EU’s lead privacy regulator ended its court proceeding related to how X processed user data to train its Grok AI chatbot, but the saga isn’t over yet for the Elon Musk-owned social media platform formerly known as Twitter. The Irish Data Protection Commission (DPC) has confirmed to TechCrunch that it’s received — and will “examine” — a number of complaints that have been filed under the bloc’s General Data Protection Regulation (GDPR).
“The DPC will now examine the extent to which any processing that has taken place complies with the relevant provisions of the GDPR,” the regulator told TechCrunch. “If, further to that examination, it is established that TUIC [Twitter International Unlimited Company, as X’s main Irish subsidiary is still known] has infringed the GDPR, the DPC will then consider whether the exercise of any of its corrective powers is warranted and, if so, which one(s).”
X agreed to suspend data processing for Grok training in early August. The undertaking X made to it was then made permanent earlier this week. That agreement committed X to delete and stop using Europeans users’ data to train its AIs which it had collected between between May 7, 2024 and August 1, 2024, according to a copy TechCrunch obtained. But it’s now clear there is no requirement on X to delete any AI models trained on the data.
So far, X hasn’t faced any sanction from the DPC for processing the personal data of Europeans to train Grok without people’s consent — despite the DPC’s urgent court action to block data collection. Penalties under the GDPR can be stiff, reaching up to 4% of global annual turnover. (Given that the company’s revenues are now in freefall, and could well scramble to hit $500 million this year based on reported quarterly figures, that might sting especially badly.)
Regulators are also empowered to order operational changes by demanding that infringement ceases. But complaints may take a long time — even up to several years — to investigate and enforce.
This is important because although X has been forced to stop helping itself to Europeans’ data to train Grok, it is still able to operate any AI models it’s already trained on the data of people who did not consent to the use — with no urgent intervention nor sanction to stop this happening as yet.
Asked whether the undertaking the DPC obtained from X last month required X to delete any AI models trained on Europeans’ data, the DPC confirmed to us it does not: “The undertaking does not require TUIC to carry out this action; it required TUIC to permanently cease the processing of the datasets covered by the undertaking,” a spokesperson said.
Some might say this is a neat way for X (or others training models) to circumvent the EU’s privacy rules: Step 1) Quietly help yourself to people’s data; Step 2) use it to train AI models and — when the cat’s out of the bag and regulators’ finally come knocking — commit to deleting *the data*, leaving your trained AI models intact. Step 3) Grok-based profit!?
Asked about this risk, the DPC responded by saying the purpose of its urgent court proceeding had been to act on “significant concerns” that X’s processing of EU and EEA users’ data to train Grok “gave rise to risk for the fundamental rights and freedoms of data subjects”. But it did not explain why it does not have the same pressing concern about risks to Europeans’ fundamental rights and freedoms from having their information embedded into Grok.
Generative AI tools are known to produce false information. Musk’s twist on the category is also intentionally irreverent — or “anti-woke” as he dubs it. That could amp up the risks about the types of content it may produce about users whose data was ingested to train the bot.
One reason the Irish regulator may be more cautious about how to deal with this issue is these AI tools are still relatively new. There’s also uncertainty among European privacy watchdogs how to enforce the GDPR against such a novel technology. Plus, it’s not clear whether the regulation’s powers would extend to being able to order AI model deletion if a technology has been trained on unlawfully processed data.
But as complaints continue to stack up in this area, data protection authorities will have to grasp the generative AI nettle sooner or later.
Going sour on Pickles
In separate X news Friday, it emerged X’s head of global affairs is out. Reuters reported the departure of long serving staffer, Nick Pickles, a U.K. national who spent a decade at Twitter, rising further up the ranks during Musk’s tenure.
In a post on X, Pickles claimed he made the decision to leave “several months ago” but does not elaborate on his reasons for leaving.
However it’s clear the company has a lot on its plate — including dealing with a ban in Brazil; and political blowback in the U.K. over its role in spreading disinformation linked to rioting in the country last month, with Musk’s personal penchant for pouring fuel on the fire (including posting on X to suggest that for the UK “civil war is inevitable”).
In the EU, X is also under investigation under the bloc’s content moderation framework. A first batch of Digital Services Act grievances were laid out in July. Musk was also recently singled out for a personal warning in an open letter penned by the bloc’s internal markets commissioner, Thierry Breton — to which the chaos-loving billionaire opted to respond with an insulting meme.
Labor shortages are still fueling growth at automation firms like GrayMatter
Robotics funding has broadly cooled off since its 2021-2022 peaks, but plenty of the issues exposed by the pandemic remain firmly in place. The biggest push behind venture funding in the category is an ongoing labor shortage. Analyst firm Garner forecasts that by 2028, half of large enterprise companies will employ robots in their warehouse and manufacturing processes.
The other key factor that warehouse and logistics robotics has going for it is a proven track record. While many approaches to automation presently have theoretical ROI, warehouse robots are out there doing the work right now, from Amazon on down.
GrayMatter is among those with a proven track record in the field. The Southern California firm self-reports that its systems currently produce “a 2-4x improvement in production line productivity [and a] 30% or more reduction in consumable waste.” Big names, including 3M, currently utilize its systems.
This is all in spite of the fact that GrayMatter is a young company, having only been founded toward the outset of the pandemic in 2020.
“We founded GrayMatter to enhance productivity while prioritizing workforce well-being,” co-founder and CEO Ariyan Kabir says in a release. “With our physics-based AI-powered systems, we are fulfilling our mission while unlocking new levels of efficiency and productivity. With our investors’ support, we are making a real difference for shop workers and addressing the critical labor shortages in manufacturing today.”
What, then, is a “physics-based” robotics system? GrayMatter contrasts its approach from the purely data-driven method used by others. The company explains:
Consider the problem of predicting process output based on the input. If the output is expected to increase with an increase in the input, then the underlying model space is limited, and a smaller amount of data can train it. We don’t need to consider arbitrarily complex models. On the other hand, this requires more complex representations and associated solution generation methods to handle constraints to produce acceptable computational performance. We cannot train a simple neural network with observed input and output data. In this case, there is no guarantee that it would preserve the process constraint if the output used during training is noisy.
Interest in the company has propelled growth. GrayMatter is a regular in our robotics job opening posts. The roundup we posted in May listed 20 open roles, among the highest of those listed.
That growth, in turn, is supported by ongoing funding. On Thursday, GrayMatter announced a $45 million Series B round, led by Wellington Management, with participation from NGP Capital, Euclidean Capital, Advance Venture Partners, SQN Venture Partners, B Capital, Bow Capital, Calibrate Ventures, OCA Ventures and Swift Ventures.
The round nearly doubles the $25 million Series A the company closed in 2022.
VCs are still pouring billions into generative AI startups
Investments in generative AI startups — those that are creating AI-powered products to generate text, audio, video and more — aren’t slowing down. But they’re being consolidated into a shrinking number of early-stage ventures.
In the first half of 2023, from January to July 16, 225 startups raised $12.3 billion from VCs, according to Crunchbase data shared with TechCrunch. Should the trend maintain, generative AI companies are on track to match or exceed the roughly $21.8 billion they raised in 2023.
Here’s how the H1 2024 total broke down by stage:
198 angel/seed deals: $500 million39 early-stage deals: $8.7 billion18 late-stage deals: $3.1 billion
Early-stage startups were the clear winners, like Elon Musk’s xAI (which raised $6 billion in May), China’s Moonshot AI ($1 billion in February), Mistral AI ($502.6 million in June), Glean ($203.2 million in February) and Cognition ($175 million in April). According to Chris Metinko, an analyst and senior reporter at Crunchbase, investors appear to be betting on big startups they see as having a high chance of success while letting those they’re less sure about “wither away” at the earlier stages.
“Some VCs expect the legal and regulatory dilemmas AI companies could face in both the U.S. and overseas to lead to a slowdown in the flood of AI funding,” Metinko told TechCrunch. “Others point to the fact that when the mobile revolution occurred more than a decade ago, the biggest winners when it came to the foundational infrastructure layer ended up being well-established tech companies.”
To Metinko’s point, the fate of many generative AI businesses — even the best-funded ones — looks murky.
Generative AI models are typically trained on data like images and text sourced from public web pages, and companies assert that fair use shields them from legal challenges in cases where that data turns out to be copyrighted. But it’s not clear yet whether the courts will ultimately decide in favor of generative AI companies, which is probably why some have begun to ink licensing deals with copyright holders.
Regardless of the outcome of any one court case, high-quality training data is becoming harder and more expensive to obtain as startups exhaust the web’s supply and more creators opt to block crawlers from scraping their data. (One analysis estimates that the market for AI training data will grow from $2.5 billion to $30 billion within a decade.) The process of training models isn’t getting any easier or cheaper, either: Per a recent Stanford report, OpenAI’s GPT-4 cost $78 million to train while Google Gemini’s price tag came in at $191 million.
Unsurprisingly given the substantial upfront investment required to build flagship models, few generative AI startups are profitable — not even big guns such as OpenAI and Anthropic. According to The Information, OpenAI, which is reportedly generating around $3.4 billion in revenue, could end up losing $5 billion this year.
Investors in generative AI are playing the long game, it’d seem — particularly big tech investors like Google, Amazon and Nvidia, which see generative AI investments as strategic bets. But could the bubble burst soon? If generative AI startups aren’t able to overcome the existential challenges facing them, that seems like a real possibility.
Tesla’s Supercharger network is still unavailable to non-Tesla EVs
It’s been more than a year since Tesla agreed to open its Supercharger network to electric vehicles from other automakers, like General Motors and Ford. But Tesla’s network of nearly 30,000 fast-charging plugs in the U.S. and Canada still remains unavailable to non-Tesla vehicles, according to a New York Times report.
The delays come amid declining sales as the automaker faces increased competition in the EV market. They also follow Tesla CEO Elon Musk’s decision earlier this year to gut the company’s Supercharger team.
Tesla posted on X last week that it had ramped production of its NACS (North American Charging Standard) adapter, which drivers of other EVs that were built with CCS (Combined Charging System) ports will need to plug into Tesla’s chargers. Still, it’s unclear how fast those adapters will make it into customers’ hands.
Will HP still demand $4B from Mike Lynch's estate?
Before entrepreneur and investor Mike Lynch died along with six others after the yacht they were on capsized in a storm last week, the party was celebrating Lynch’s victory in the U.S. criminal courts. In June, he was acquitted of all counts of fraud connected to HP’s 2011 acquisition of his company, Autonomy. But it was not the final chapter in that dispute. HP (now known as HPE) was still trying to recover $4 billion from him as a result of a civil case Lynch lost in the U.K.
Now HPE faces the question: Does it forge ahead, even with Lynch deceased?
An article in Fortune notes HPE could be facing a PR disaster if it does; further, the estate is unlikely to be able to do anything until it resolves any other litigation or appeals. Still, $4 billion is no small sum, leading a legal expert to colorfully describe HPE’s predicament to the outlet as “on the horns of a dilemma.”
You can read more about HPE’s options here.
Labor shortages are still fueling growth at automation firms like GrayMatter
Robotics funding has broadly cooled off since its 2021-2022 peaks, but plenty of the issues exposed by the pandemic remain firmly in place. The biggest push behind venture funding in the category is an ongoing labor shortage. Analyst firm Garner forecasts that by 2028, half of large enterprise companies will employ robots in their warehouse and manufacturing processes.
The other key factor that warehouse and logistics robotics has going for it is a proven track record. While many approaches to automation presently have theoretical ROI, warehouse robots are out there doing the work right now, from Amazon on down.
GrayMatter is among those with a proven track record in the field. The Southern California firm self-reports that its systems currently produce “a 2-4x improvement in production line productivity [and a] 30% or more reduction in consumable waste.” Big names, including 3M, currently utilize its systems.
This is all in spite of the fact that GrayMatter is a young company, having only been founded toward the outset of the pandemic in 2020.
“We founded GrayMatter to enhance productivity while prioritizing workforce well-being,” co-founder and CEO Ariyan Kabir says in a release. “With our physics-based AI-powered systems, we are fulfilling our mission while unlocking new levels of efficiency and productivity. With our investors’ support, we are making a real difference for shop workers and addressing the critical labor shortages in manufacturing today.”
What, then, is a “physics-based” robotics system? GrayMatter contrasts its approach from the purely data-driven method used by others. The company explains:
Consider the problem of predicting process output based on the input. If the output is expected to increase with an increase in the input, then the underlying model space is limited, and a smaller amount of data can train it. We don’t need to consider arbitrarily complex models. On the other hand, this requires more complex representations and associated solution generation methods to handle constraints to produce acceptable computational performance. We cannot train a simple neural network with observed input and output data. In this case, there is no guarantee that it would preserve the process constraint if the output used during training is noisy.
Interest in the company has propelled growth. GrayMatter is a regular in our robotics job opening posts. The roundup we posted in May listed 20 open roles, among the highest of those listed.
That growth, in turn, is supported by ongoing funding. On Thursday, GrayMatter announced a $45 million Series B round, led by Wellington Management, with participation from NGP Capital, Euclidean Capital, Advance Venture Partners, SQN Venture Partners, B Capital, Bow Capital, Calibrate Ventures, OCA Ventures and Swift Ventures.
The round nearly doubles the $25 million Series A the company closed in 2022.
VCs are still pouring billions into generative AI startups
Investments in generative AI startups — those that are creating AI-powered products to generate text, audio, video and more — aren’t slowing down. But they’re being consolidated into a shrinking number of early-stage ventures.
In the first half of 2023, from January to July 16, 225 startups raised $12.3 billion from VCs, according to Crunchbase data shared with TechCrunch. Should the trend maintain, generative AI companies are on track to match or exceed the roughly $21.8 billion they raised in 2023.
Here’s how the H1 2024 total broke down by stage:
198 angel/seed deals: $500 million39 early-stage deals: $8.7 billion18 late-stage deals: $3.1 billion
Early-stage startups were the clear winners, like Elon Musk’s xAI (which raised $6 billion in May), China’s Moonshot AI ($1 billion in February), Mistral AI ($502.6 million in June), Glean ($203.2 million in February) and Cognition ($175 million in April). According to Chris Metinko, an analyst and senior reporter at Crunchbase, investors appear to be betting on big startups they see as having a high chance of success while letting those they’re less sure about “wither away” at the earlier stages.
“Some VCs expect the legal and regulatory dilemmas AI companies could face in both the U.S. and overseas to lead to a slowdown in the flood of AI funding,” Metinko told TechCrunch. “Others point to the fact that when the mobile revolution occurred more than a decade ago, the biggest winners when it came to the foundational infrastructure layer ended up being well-established tech companies.”
To Metinko’s point, the fate of many generative AI businesses — even the best-funded ones — looks murky.
Generative AI models are typically trained on data like images and text sourced from public web pages, and companies assert that fair use shields them from legal challenges in cases where that data turns out to be copyrighted. But it’s not clear yet whether the courts will ultimately decide in favor of generative AI companies, which is probably why some have begun to ink licensing deals with copyright holders.
Regardless of the outcome of any one court case, high-quality training data is becoming harder and more expensive to obtain as startups exhaust the web’s supply and more creators opt to block crawlers from scraping their data. (One analysis estimates that the market for AI training data will grow from $2.5 billion to $30 billion within a decade.) The process of training models isn’t getting any easier or cheaper, either: Per a recent Stanford report, OpenAI’s GPT-4 cost $78 million to train while Google Gemini’s price tag came in at $191 million.
Unsurprisingly given the substantial upfront investment required to build flagship models, few generative AI startups are profitable — not even big guns such as OpenAI and Anthropic. According to The Information, OpenAI, which is reportedly generating around $3.4 billion in revenue, could end up losing $5 billion this year.
Investors in generative AI are playing the long game, it’d seem — particularly big tech investors like Google, Amazon and Nvidia, which see generative AI investments as strategic bets. But could the bubble burst soon? If generative AI startups aren’t able to overcome the existential challenges facing them, that seems like a real possibility.
VCs are still pouring billions into generative AI startups
Investments in generative AI startups — those that are creating AI-powered products to generate text, audio, video and more — aren’t slowing down. But they’re being consolidated into a shrinking number of early-stage ventures.
In the first half of 2023, from January to July 16, 225 startups raised $12.3 billion from VCs, according to Crunchbase data shared with TechCrunch. Should the trend maintain, generative AI companies are on track to match or exceed the roughly $21.8 billion they raised in 2023.
Here’s how the H1 2024 total broke down by stage:
198 angel/seed deals: $500 million39 early-stage deals: $8.7 billion18 late-stage deals: $3.1 billion
Early-stage startups were the clear winners, like Elon Musk’s xAI (which raised $6 billion in May), China’s Moonshot AI ($1 billion in February), Mistral AI ($502.6 million in June), Glean ($203.2 million in February) and Cognition ($175 million in April). According to Chris Metinko, an analyst and senior reporter at Crunchbase, investors appear to be betting on big startups they see as having a high chance of success while letting those they’re less sure about “wither away” at the earlier stages.
“Some VCs expect the legal and regulatory dilemmas AI companies could face in both the U.S. and overseas to lead to a slowdown in the flood of AI funding,” Metinko told TechCrunch. “Others point to the fact that when the mobile revolution occurred more than a decade ago, the biggest winners when it came to the foundational infrastructure layer ended up being well-established tech companies.”
To Metinko’s point, the fate of many generative AI businesses — even the best-funded ones — looks murky.
Generative AI models are typically trained on data like images and text sourced from public web pages, and companies assert that fair use shields them from legal challenges in cases where that data turns out to be copyrighted. But it’s not clear yet whether the courts will ultimately decide in favor of generative AI companies, which is probably why some have begun to ink licensing deals with copyright holders.
Regardless of the outcome of any one court case, high-quality training data is becoming harder and more expensive to obtain as startups exhaust the web’s supply and more creators opt to block crawlers from scraping their data. (One analysis estimates that the market for AI training data will grow from $2.5 billion to $30 billion within a decade.) The process of training models isn’t getting any easier or cheaper, either: Per a recent Stanford report, OpenAI’s GPT-4 cost $78 million to train while Google Gemini’s price tag came in at $191 million.
Unsurprisingly given the substantial upfront investment required to build flagship models, few generative AI startups are profitable — not even big guns such as OpenAI and Anthropic. According to The Information, OpenAI, which is reportedly generating around $3.4 billion in revenue, could end up losing $5 billion this year.
Investors in generative AI are playing the long game, it’d seem — particularly big tech investors like Google, Amazon and Nvidia, which see generative AI investments as strategic bets. But could the bubble burst soon? If generative AI startups aren’t able to overcome the existential challenges facing them, that seems like a real possibility.