The first self-checkout system was installed in 1986 in a Kroger grocery store just outside of Atlanta. It took several decades, but the technology has finally proliferated across the U.S. Given the automated direction grocery stores are heading, it seems that robotic bagging can’t be too far behind.
MIT’s CSAIL department this week is showcasing RoboGrocery. It combines computer vision with a soft robotic gripper to bag a wide range of items. To test the system, researchers placed 10 objects unknown to the robot on a grocery conveyer belt.
The products ranged from delicate items like grapes, bread, kale, muffins and crackers to far more solid ones like soup cans, meal boxes and ice cream containers. The vision system kicks in first, detecting the objects before determining their size and orientation on the belt.
As the grasper touches the grapes, pressure sensors in the fingers determine that they are, in fact, delicate and therefore should not go at the bottom of the bag — something many of us no doubt learned the hard way. Next, it notes that the soup can is a more rigid structure and sticks it in the bottom of the bag.
“This is a significant first step towards having robots pack groceries and other items in real-world settings,” said Annan Zhang, one of the study’s lead authors. “Although we’re not quite ready for commercial deployment, our research demonstrates the power of integrating multiple sensing modalities in soft robotic systems.”
The team notes that there’s still plenty of room for improvement, including upgrades to the grasper and the imaging system to better determine how and in what order to pack things. As the system becomes more robust, it may also be scaled outside the grocery into more industrial spaces like recycling plants.
Generative AI models aren’t actually humanlike. They have no intelligence or personality — they’re simply statistical systems predicting the likeliest next words in a sentence. But like interns at a tyrannical workplace, they do follow instructions without complaint — including initial “system prompts” that prime the models with their basic qualities and what they should and shouldn’t do.
Every generative AI vendor, from OpenAI to Anthropic, uses system prompts to prevent (or at least try to prevent) models from behaving badly, and to steer the general tone and sentiment of the models’ replies. For instance, a prompt might tell a model it should be polite but never apologetic, or to be honest about the fact that it can’t know everything.
But vendors usually keep system prompts close to the chest — presumably for competitive reasons, but also perhaps because knowing the system prompt may suggest ways to circumvent it. The only way to expose GPT-4o‘s system prompt, for example, is through a prompt injection attack. And even then, the system’s output can’t be trusted completely.
However, Anthropic, in its continued effort to paint itself as a more ethical, transparent AI vendor, has published the system prompts for its latest models (Claude 3 Opus, Claude 3.5 Sonnet and Claude 3 Haiku) in the Claude iOS and Android apps and on the web.
Alex Albert, head of Anthropic’s developer relations, said in a post on X that Anthropic plans to make this sort of disclosure a regular thing as it updates and fine-tunes its system prompts.
We've added a new system prompts release notes section to our docs. We're going to log changes we make to the default system prompts on Claude dot ai and our mobile apps. (The system prompt does not affect the API.) pic.twitter.com/9mBwv2SgB1
— Alex Albert (@alexalbert__) August 26, 2024
The latest prompts, dated July 12, outline very clearly what the Claude models can’t do — e.g. “Claude cannot open URLs, links, or videos.” Facial recognition is a big no-no; the system prompt for Claude Opus tells the model to “always respond as if it is completely face blind” and to “avoid identifying or naming any humans in [images].”
But the prompts also describe certain personality traits and characteristics — traits and characteristics that Anthropic would have the Claude models exemplify.
The prompt for Claude 3 Opus, for instance, says that Claude is to appear as if it “[is] very smart and intellectually curious,” and “enjoys hearing what humans think on an issue and engaging in discussion on a wide variety of topics.” It also instructs Claude to treat controversial topics with impartiality and objectivity, providing “careful thoughts” and “clear information” — and never to begin responses with the words “certainly” or “absolutely.”
It’s all a bit strange to this human, these system prompts, which are written like an actor in a stage play might write a character analysis sheet. The prompt for Opus ends with “Claude is now being connected with a human,” which gives the impression that Claude is some sort of consciousness on the other end of the screen whose only purpose is to fulfill the whims of its human conversation partners.
But of course that’s an illusion. If the prompts for Claude tell us anything, it’s that without human guidance and hand-holding, these models are frighteningly blank slates.
With these new system prompt changelogs — the first of their kind from a major AI vendor — Anthropic is exerting pressure on competitors to publish the same. We’ll have to see if the gambit works.
Generative AI models aren’t actually humanlike. They have no intelligence or personality — they’re simply statistical systems predicting the likeliest next words in a sentence. But like interns at a tyrannical workplace, they do follow instructions without complaint — including initial “system prompts” that prime the models with their basic qualities and what they should and shouldn’t do.
Every generative AI vendor, from OpenAI to Anthropic, uses system prompts to prevent (or at least try to prevent) models from behaving badly, and to steer the general tone and sentiment of the models’ replies. For instance, a prompt might tell a model it should be polite but never apologetic, or to be honest about the fact that it can’t know everything.
But vendors usually keep system prompts close to the chest — presumably for competitive reasons, but also perhaps because knowing the system prompt may suggest ways to circumvent it. The only way to expose GPT-4o‘s system prompt, for example, is through a prompt injection attack. And even then, the system’s output can’t be trusted completely.
However, Anthropic, in its continued effort to paint itself as a more ethical, transparent AI vendor, has published the system prompts for its latest models (Claude 3 Opus, Claude 3.5 Sonnet and Claude 3 Haiku) in the Claude iOS and Android apps and on the web.
Alex Albert, head of Anthropic’s developer relations, said in a post on X that Anthropic plans to make this sort of disclosure a regular thing as it updates and fine-tunes its system prompts.
We've added a new system prompts release notes section to our docs. We're going to log changes we make to the default system prompts on Claude dot ai and our mobile apps. (The system prompt does not affect the API.) pic.twitter.com/9mBwv2SgB1
— Alex Albert (@alexalbert__) August 26, 2024
The latest prompts, dated July 12, outline very clearly what the Claude models can’t do — e.g. “Claude cannot open URLs, links, or videos.” Facial recognition is a big no-no; the system prompt for Claude Opus tells the model to “always respond as if it is completely face blind” and to “avoid identifying or naming any humans in [images].”
But the prompts also describe certain personality traits and characteristics — traits and characteristics that Anthropic would have the Claude models exemplify.
The prompt for Claude 3 Opus, for instance, says that Claude is to appear as if it “[is] very smart and intellectually curious,” and “enjoys hearing what humans think on an issue and engaging in discussion on a wide variety of topics.” It also instructs Claude to treat controversial topics with impartiality and objectivity, providing “careful thoughts” and “clear information” — and never to begin responses with the words “certainly” or “absolutely.”
It’s all a bit strange to this human, these system prompts, which are written like an actor in a stage play might write a character analysis sheet. The prompt for Opus ends with “Claude is now being connected with a human,” which gives the impression that Claude is some sort of consciousness on the other end of the screen whose only purpose is to fulfill the whims of its human conversation partners.
But of course that’s an illusion. If the prompts for Claude tell us anything, it’s that without human guidance and hand-holding, these models are frighteningly blank slates.
With these new system prompt changelogs — the first of their kind from a major AI vendor — Anthropic is exerting pressure on competitors to publish the same. We’ll have to see if the gambit works.
Website building platform Squarespace is selling Tock, its restaurant reservation service, to American Express in a deal worth $400 million — the exact figure that Squarespace paid for the service three years ago. Separately, AmEx will be buying Rooam, a mobile payments and ordering platform that integrates with point-of-sale systems.
Terms of the Rooam deal were not disclosed.
The news comes a month after private equity giant Permira announced plans to acquire Squarespace and take it private in a deal worth $6.9 billion. That transaction is expected to close in Q4 2024.
Founded in 2014, Tock targets the hospitality industry with reservation and event management tools, with a particular focus on restaurants looking to integrate table management and booking services, or sell tickets for special events such as wine tastings.
Squarespace acquired Tock in 2021 for $400 million, shortly before going public on the New York Stock Exchange (NYSE). The move was somewhat synergistic to its main website building offering, insofar as Square could offer restaurant customers a way not only to develop their online visibility, but then build further services around that.
That Squarespace is divesting Tock before its own acquisition is concluded suggests that it sees Tock as surplus to requirements regardless of the outcome, and an extra $400 million in the bank will enable it to get back to its roots as a website builder for SMBs, while investing in other complementary tools such as online payments for e-commerce firms.
TechCrunch has reached out both to Permira and Squarespace to get further insights on the rationale behind this sale, and will update when we hear back.
The first self-checkout system was installed in 1986 in a Kroger grocery store just outside of Atlanta. It took several decades, but the technology has finally proliferated across the U.S. Given the automated direction grocery stores are heading, it seems that robotic bagging can’t be too far behind.
MIT’s CSAIL department this week is showcasing RoboGrocery. It combines computer vision with a soft robotic gripper to bag a wide range of items. To test the system, researchers placed 10 objects unknown to the robot on a grocery conveyer belt.
The products ranged from delicate items like grapes, bread, kale, muffins and crackers to far more solid ones like soup cans, meal boxes and ice cream containers. The vision system kicks in first, detecting the objects before determining their size and orientation on the belt.
As the grasper touches the grapes, pressure sensors in the fingers determine that they are, in fact, delicate and therefore should not go at the bottom of the bag — something many of us no doubt learned the hard way. Next, it notes that the soup can is a more rigid structure and sticks it in the bottom of the bag.
“This is a significant first step towards having robots pack groceries and other items in real-world settings,” said Annan Zhang, one of the study’s lead authors. “Although we’re not quite ready for commercial deployment, our research demonstrates the power of integrating multiple sensing modalities in soft robotic systems.”
The team notes that there’s still plenty of room for improvement, including upgrades to the grasper and the imaging system to better determine how and in what order to pack things. As the system becomes more robust, it may also be scaled outside the grocery into more industrial spaces like recycling plants.
The first self-checkout system was installed in 1986 in a Kroger grocery store just outside of Atlanta. It took several decades, but the technology has finally proliferated across the U.S. Given the automated direction grocery stores are heading, it seems that robotic bagging can’t be too far behind.
MIT’s CSAIL department this week is showcasing RoboGrocery. It combines computer vision with a soft robotic gripper to bag a wide range of items. To test the system, researchers placed 10 objects unknown to the robot on a grocery conveyer belt.
The products ranged from delicate items like grapes, bread, kale, muffins and crackers to far more solid ones like soup cans, meal boxes and ice cream containers. The vision system kicks in first, detecting the objects before determining their size and orientation on the belt.
As the grasper touches the grapes, pressure sensors in the fingers determine that they are, in fact, delicate and therefore should not go at the bottom of the bag — something many of us no doubt learned the hard way. Next, it notes that the soup can is a more rigid structure and sticks it in the bottom of the bag.
“This is a significant first step towards having robots pack groceries and other items in real-world settings,” said Annan Zhang, one of the study’s lead authors. “Although we’re not quite ready for commercial deployment, our research demonstrates the power of integrating multiple sensing modalities in soft robotic systems.”
The team notes that there’s still plenty of room for improvement, including upgrades to the grasper and the imaging system to better determine how and in what order to pack things. As the system becomes more robust, it may also be scaled outside the grocery into more industrial spaces like recycling plants.
Website building platform Squarespace is selling Tock, its restaurant reservation service, to American Express in a deal worth $400 million — the exact figure that Squarespace paid for the service three years ago. Separately, AmEx will be buying Rooam, a mobile payments and ordering platform that integrates with point-of-sale systems.
Terms of the Rooam deal were not disclosed.
The news comes a month after private equity giant Permira announced plans to acquire Squarespace and take it private in a deal worth $6.9 billion. That transaction is expected to close in Q4 2024.
Founded in 2014, Tock targets the hospitality industry with reservation and event management tools, with a particular focus on restaurants looking to integrate table management and booking services, or sell tickets for special events such as wine tastings.
Squarespace acquired Tock in 2021 for $400 million, shortly before going public on the New York Stock Exchange (NYSE). The move was somewhat synergistic to its main website building offering, insofar as Square could offer restaurant customers a way not only to develop their online visibility, but then build further services around that.
That Squarespace is divesting Tock before its own acquisition is concluded suggests that it sees Tock as surplus to requirements regardless of the outcome, and an extra $400 million in the bank will enable it to get back to its roots as a website builder for SMBs, while investing in other complementary tools such as online payments for e-commerce firms.
TechCrunch has reached out both to Permira and Squarespace to get further insights on the rationale behind this sale, and will update when we hear back.
Federal safety regulators have opened an investigation into Ford’s hands-free driver-assistance system, BlueCruise, after it was found to be active during two recent crashes that killed multiple people.
The National Highway Traffic Safety Administration’s Office of Defects Investigation (ODI) said on Monday that it has confirmed BlueCruise was active in both crashes. One of the crashes happened in February in Texas, and the other occurred in early April in Pennsylvania. They are the first known fatalities resulting from crashes involving the use of BlueCruise.
The investigation into the two crashes ratchets up the scrutiny on BlueCruise, which is currently available on the Mustang Mach-E, and certain Ford F-150s (including the Lightning), Explorers and Expeditions. The National Transportation Safety Board has already opened an investigation into the Texas crash. Ford said in a statement that it is “working with NHTSA to support its investigation.”
The new probe comes just days after ODI closed its most high-profile driver-assistance investigation to date. The safety agency initially opened an investigation into Tesla’s Autopilot in 2021 after multiple reports of Teslas crashing into stationary emergency vehicles while the drivers were using Autopilot. In closing the investigation, the ODI said last week that it determined a “critical safety gap between drivers’ expectations of [Autopilot’s] operating capabilities and the system’s true capabilities” created “foreseeable misuse and avoidable crashes.”
Ford announced BlueCruise in 2021. It is only available on pre-mapped highways, and Ford pairs it with a camera-based driver monitoring system that checks whether drivers’ eyes are still on the road when the system is active. Those represent much tighter constraints on the system than Tesla puts on the use of Autopilot.
Still, while BlueCruise is highly rated by some, including Consumer Reports, the recent crashes and resulting investigations suggest there may be a more fundamental problem with advanced driver-assistance systems than some of these companies are willing to admit.
Note: This story has been updated to include a response from Ford.
Lizcore, a sport tracking startup out of Barcelona, caught our eye on the 4YFN show floor at MWC this week. It’s done what this climber — frankly — thought impossible: Come up with a way to digitize and upgrade the indoor climbing experience that bouldering addicts might actually get on board with.
Bouldering gyms that install its proprietary blend of hardware and software can offer climbers a low friction way to log and track their indoor climbing performance (recording on-sights, red points, how fast problems were scaled, etc.) — while also adding a dash of social gamification (such as the ability to compare stats with friends) — all without having to wear or carry anything more arduous than a slender fabric bracelet or clothes pin.
“Every sport is going from analogue to digital. And climbing was an exception in that sense,” says co-founder and CTO Marçal Juan. “The main issue is because you cannot create an app and expect that climbers use it. So that’s why we need some hardware and that’s why we can get success with that.”
To make use of Lizcore’s progress tracking system, climbers don’t need to climb with their phone or wear anything as annoyingly chunky as a smartwatch (such wearables can just get in the way of climbing and risk being scratched and soiled in a chalky gym). All they need is its lightweight NFC wearable (aka the Lizy bracelet). This minimalist band comes in a range of snazzy colors and is the only bit of hardware required to pair with the startup’s app.
The team is also working on designing an even less intrusive NFC clothes pin, too, to offer a second climber-friendly alternative to persuade boulders this tech won’t weigh them down.
Lizcore’s system requires gyms to buy in and install its smart base units and top-out holds for each boulder problem. One base unit can be shared by up to three routes. Climbers identify themselves at the base unit, using the NFC wearable. Then they make their attempt at the route and the app records either a fail if they don’t reach the top hold or a send and a time for topping out if they make it.
It’s activity and progress tracking with minimal effort so climbers can concentrate on the real work — of training, climbing, sending and repeating. Unsurprisingly the idea for the business came from an ex-pro climber: Co-founder and CEO, Edgar Casanovas.
The wall-mounted Lizcore base unit features a screen so it can display grading info per route. (When we visited the stand the screen wasn’t displaying, owing, we were told, to issues with the conference Wi-Fi.) The buttons on the unit can also be used for climbers to vote on whether or not they agree with the grade after the fact. The unit can also provide them with the ability to vote for a route to be maintained for longer (instead of being reset with a new problem, as bouldering gyms routinely do).
Lizcore says its NFC system could also work for gyms to run ticketing at the entrance/exit, if they don’t already have a system installed, and/or to provide access to fitness classes they may also provide. Another possible use is security — to restrict entrance to particular sections of the gym (such as to prevent kids using certain facilities). And, being NFC, the lightweight devices could support payments, such as at the café or bar gyms typically run. So climbers could even use the minimalist wearable to forget having to bring their own wallets to the gym.
Lizcore CTO Marçal Juan shows off the Lizy band. Image Credits: Natasha Lomas/TechCrunch
“The possibilities are endless,” suggests Juan. “You can expect that the [gyms which] buy this device want to fully optimise the usage. So if we sell another service, it’s very easy to say, okay, I’m interested. The selling point is that we have a hardware that is needed for climbing. And also, because we can have a lot of services… the CRM, the subscriptions, you can use it for everything.”
The two-sided system is also designed to serve up plenty of data insights for gyms too, providing these businesses with analytics and a more granular overview of climbing activity than if they’re just tracking gymgoers’ comings and goings via an entry/exit gate. Lizcore says the system gives gyms data about quiet versus busy times for different climbing areas and walls. Or even how specific routes are performing (too easy, too hard, etc.) — to help them balance the mix of problems offered.
It also provides a way for gyms to run competitions for users — for example, Juan says they could run prize giveaways for the first or fastest ascent of new boulder problems. Or use the tracked routes to verify if climbers have successfully topped out during open bouldering competitions.
With the app sitting on climbers’ phones it opens up a mobile channel for gyms to more easily reach their core users, say with notifications about offers and promotions. But also to offer members more value: Such as providing info on new routes as soon as they’re set and detailed per-route info.
Gyms often rely on posting to Instagram with these sorts of ‘new routes’ PSAs but, via Lizcore, they could target members with notifications when routes have been put up and provide them with the ability to check out route setting in the app ahead of a visit, to read problems and better plan and prepare their trips to the gym.
More information could also be provided by route setters — such as whether a particular route needs a lot of finger strength, power or dynamic technique.
Normally, in bouldering gyms, the only info climbers get is much more basic: A route grade (or, often, just a route range). Add to that, a route grade may vary depending on the morphology of the individual climber (small vs tall for example) — so Lizcore’s route tracking and voting system could also offer a way for climbers to be served more individualized route grades. (Speaking as a short climber that would really be a data-driven leap forward!)
The November 2022-founded startup is working with a handful of gyms (five around Spain) on early deployments of its system so far. It’s currently looking to raise a seed round to switch into a mass production mode, per Juan, who says they’re focusing on scaling in Spain and Europe initially.
He says they’re also in talks with climbing competition organizations about using their system, which also supports uploading audio route flash descriptions to the app for visually impaired climbers.
Additionally, the team is working on developing safety tech for climbing gyms — such as smart auto belay systems that won’t let a person clip into the carabiner unless they have the necessary credential to use the equipment.
Image Credits: Snap CEO Evan Spiegel; / TechCrunch
Less than a week after The Wall St. Journal reported on how a Snapchat feature dubbed “solar system” was adding to teens’ anxiety, the company has responded by adjusting how the feature works. The ranking system for paid subscribers today shows you how close you are to your Snapchat friends by displaying your position in their solar system. For example, a friend in the “Mercury” position would be someone you communicate with a lot, while “Uranus” would be someone not as close.
Of course, online chatting doesn’t necessarily correlate to real-world relationships, and such a feature can lead to hurt feelings when someone realizes that they’re not as close to a friend as they thought.
Snap says it has received feedback that it can feel good to know you’re close to someone but it can also feel bad to know you aren’t as close as you’d like to be.
“We’ve heard and understand that the Solar System can make that feeling worse, and we want to avoid that,” the company announced in a post on Friday.
However, instead of removing the feature, as it did with the dangerous and controversial speed filter, which it was sued over for “negligent design,” Snap is simply turning the Solar System feature off by default. Snapchat+ subscribers will still be able to turn the option on if they choose.
“We hope this strikes the right balance between providing a feature that is desired by many who use it while avoiding upsetting those who don’t want to use it,” the company explains.
Turning it off by default may provide some friction, but if the feature is already in demand among teens, then they’ll simply dig around to find the setting to turn it back on.
Snap argues that Solar System is not that popular, noting that less than 0.25% of the community uses the option. But since it’s only available to paid subscribers, the small percentage is not surprising. A more relevant stat would be how many Snapchat+ users have used Solar System or viewed the feature.
Although users can’t see who’s closer or farther away from the friend as they are, finding out they’re not number one has led to some tough conversations, The WSJ reported, even breakups.
Snap defends the feature by saying that people wanted to know more about their friendships, and features like Solar System provide “additional awareness and context.” But in reality, it’s a way to keep young people — a demographic where social hierarchy is key — addicted to using Snapchat.
The Solar System feature was only one of Snapchat’s friend ranking systems. It also offers a private feature called “Best Friends” that puts the people with whom you communicate most at the top of your contact list, along with a heart or smiley emoji, The WSJ pointed out.
Another much-debated feature called “Streaks” is a tool that Snapchat uses to encourage repeated use of its app by offering a visual representation of how many consecutive days users have stayed in touch with one another on the app. After much backlash from parents and families, lawmakers, and regulators alike over the feature’s addictive nature and psychological harms, Snap last year introduced a way to pause your streaks. It also added a way for users to restore a lost Streak.
While Snap promises in its blog post that it’s “committed to mitigating the potential downsides of online communication wherever possible,” it has intentionally built features and tools that have at least left it open to lawsuits and Congressional inquiry, if not worse.
CEOs from Meta, TikTok, Snap, X and Discord head to Congress for kids’ online safety hearing
YouTube and Snapchat were asked to defend their apps’ age ratings in Senate hearing
