SAN FRANCISCO (Reuters) - Apple Inc (AAPL.O) dropped plans to let iPhone users fully encrypt backups of their devices in the company’s iCloud service after the FBI complained that the move would harm investigations, six sources familiar with the matter told Reuters.
FILE PHOTO: A woman uses her Apple iPhone and laptop in a cafe in lower Manhattan in New York City, U.S., May 8, 2019. REUTERS/Mike Segar/File Photo
The tech giant’s reversal, about two years ago, has not previously been reported. It shows how much Apple has been willing to help U.S. law enforcement and intelligence agencies, despite taking a harder line in high-profile legal disputes with the government and casting itself as a defender of its customers’ information.
The long-running tug of war between investigators’ concerns about security and tech companies’ desire for user privacy moved back into the public spotlight last week, as U.S. Attorney General William Barr took the rare step of publicly calling on Apple to unlock two iPhones used by a Saudi Air Force officer who shot dead three Americans at a Pensacola, Florida naval base last month.
U.S. President Donald Trump piled on, accusing Apple on Twitter of refusing to unlock phones used by “killers, drug dealers and other violent criminal elements.” Republican and Democratic senators sounded a similar theme in a December hearing, threatening legislation against end-to-end encryption, citing unrecoverable evidence of crimes against children.
Apple did in fact did turn over the shooter’s iCloud backups in the Pensacola case, and said it rejected the characterization that it “has not provided substantive assistance.”
Behind the scenes, Apple has provided the U.S. Federal Bureau of Investigation with more sweeping help, not related to any specific probe.
An Apple spokesman declined to comment on the company’s handling of the encryption issue or any discussions it has had with the FBI. The FBI did not respond to requests for comment on any discussions with Apple.
More than two years ago, Apple told the FBI that it planned to offer users end-to-end encryption when storing their phone data on iCloud, according to one current and three former FBI officials and one current and one former Apple employee.
Under that plan, primarily designed to thwart hackers, Apple would no longer have a key to unlock the encrypted data, meaning it would not be able to turn material over to authorities in a readable form even under court order.
In private talks with Apple soon after, representatives of the FBI’s cyber crime agents and its operational technology division objected to the plan, arguing it would deny them the most effective means for gaining evidence against iPhone-using suspects, the government sources said.
When Apple spoke privately to the FBI about its work on phone security the following year, the end-to-end encryption plan had been dropped, according to the six sources. Reuters could not determine why exactly Apple dropped the plan.
“Legal killed it, for reasons you can imagine,” another former Apple employee said he was told, without any specific mention of why the plan was dropped or if the FBI was a factor in the decision.
That person told Reuters the company did not want to risk being attacked by public officials for protecting criminals, sued for moving previously accessible data out of reach of government agencies or used as an excuse for new legislation against encryption.
“They decided they weren’t going to poke the bear anymore,” the person said, referring to Apple’s court battle with the FBI in 2016 over access to an iPhone used by one of the suspects in a mass shooting in San Bernardino, California.
Apple appealed a court order to break into that phone for the FBI. The government dropped the proceedings when it found a contractor that could break into the phone, a common occurrence in FBI investigations.
Two of the former FBI officials, who were not present in talks with Apple, told Reuters it appeared that the FBI’s arguments that the backups provided vital evidence in thousands of cases had prevailed.
“It’s because Apple was convinced,” said one. “Outside of that public spat over San Bernardino, Apple gets along with the federal government.”
However, a former Apple employee said it was possible the encryption project was dropped for other reasons, such as concern that more customers would find themselves locked out of their data more often.
Once the decision was made, the 10 or so experts on the Apple encryption project - variously code-named Plesio and KeyDrop - were told to stop working on the effort, three people familiar with the matter told Reuters.
APPLE SHIFTS FOCUS
Apple’s decision not to proceed with end-to-end encryption of iCloud backups made the FBI’s job easier.
The agency relies on hacking software that exploits security flaws to break into a phone. But that method requires direct access to the phone which would ordinarily tip off the user, who is often the subject of the investigation.
Apple’s iCloud, on the other hand, can be searched in secret. In the first half of last year, the period covered by Apple’s most recent semiannual transparency report on requests for data it receives from government agencies, U.S. authorities armed with regular court papers asked for and obtained full device backups or other iCloud content in 1,568 cases, covering about 6,000 accounts.
The company said it turned over at least some data for 90% of the requests it received. It turns over data more often in response to secret U.S. intelligence court directives, topping 14,000 accounts in the second half of 2018. Because of gag orders, Apple has not given any such data for 2019.
Had it proceeded with its plan, Apple would not have been able to turn over any readable data belonging to users who opted for end-to-end encryption.
Instead of protecting all of iCloud with end-to-end encryption, Apple has shifted to focus on protecting some of the most sensitive user information, such as saved passwords and health data.
But backed-up contact information and texts from iMessage, WhatsApp and other encrypted services remain available to Apple employees and authorities.
Apple is not the only tech company to have removed its own access to customers’ information.
Slideshow (2 Images)
In October 2018, Alphabet Inc’s (GOOGL.O) Google announced a similar system to Apple’s dropped plan for secure backups. The maker of Android software, which runs on about three-quarters of the world’s mobile devices, said users could back up their data to its own cloud without trusting the company with the key.
Two people familiar with the project said Google gave no advance notice to governments, and picked a time to announce it when encryption was not in the news.
The company continues to offer the service but declined to comment on how many users have taken up the option. The FBI did not respond to a request for comment on Google’s service or the agency’s approach to it.
Reporting by Joseph Menn in San Francisco; Editing by Bill Rigby
I couldn't be more excited to see Samsung's next foldable phone. At its Unpacked event on Feb. 11, the world's largest phone maker is expected to show a design that flips up like the Motorola Razr instead of opening into a small tablet like the previous Galaxy Fold. Rumor says it'll be commercially called the Galaxy Z Flip. Internal code-name, Galaxy Bloom.
There's something about foldable phones that still feels magical to me, a sensation that's hard to come by after nearly 14 years working with phones. I thought I'd seen it all: phones that flip up, kick out, flex, even phones shaped like a circle. But a screen that's both thin and strong enough to bend in half without breaking feels like the future come to life.
For Samsung, the Z Flip will give the company a dramatic lead in the foldable space, with a large premium device (the Galaxy Fold) followed by a simpler phone with a smaller screen. That's two pathways for Samsung to secure interest from early adopters.
For the rest of us, there are several practical reasons to train our eyes on the Galaxy Z Flip (or whatever it winds up being called) that go way beyond the novelty of a foldable phone.
An artistic render of how the Galaxy Z Flip could look.
Giuseppe Spinelli/LetsGoDigital
The first foldable phone with a glass screen?
One rumor suggests that the Galaxy Z Flip will be the first foldable design to use an ultrathin glass screen. That's momentous if true. The first generation of foldables use plastic displays, which are much more susceptible to damage.
It'd be interesting to see how completely a glass screen would bend. Could the two sides really fold flat? I'd also love to test firsthand the screen's ability to keep the electronic display underneath safe from pressure, scratches, drops, dust and water damage.
Now playing:Watch this:
The bendable glass that’s shaping up to cover foldable...
4:06
Foldable phones are still a proof of concept
Right now, foldable phones inhabit a zone of uncertainty. They're expensive, fragile and few. At this point it's hard to believe that they could replace the large-screen rectangles we carry today, but there are hints it just might work.
The more foldable phones exist -- both in design and in total number of units made -- the more we can see if they'll actually take off. Or if they're just fun, expensive toys. The Z Flip will be one more effort that helps determine the fate of the category.
The variety we see already in early foldable designs is crucial. We've seen commercial devices and prototypes for small foldables like the Razr, which can slip into a pocket, all the way up to a 10-inch tablet that folds into three parts.
It will be through real-time trial and error that the industry determines which designs work best, how to fix common weaknesses and what it is that people actually want in a foldable phone. Only then can companies collectively begin to perfect them.
This is Samsung's chance to prove it can make a sturdy foldable
An embarrassment for Samsung, the Galaxy Fold's early screen flaws overshadowed its historic debut. Samsung delayed the initial sale date for months, redesigned the phone, scaled back production and dropped two colors. Now with the rumored Galaxy Z Flip, Samsung gets a second chance.
Choosing a radically different design -- the Galaxy Z Flip should be a vertical flip phone with a smaller screen than the 7.3-inch Galaxy Fold -- gives Samsung an opportunity to apply the lessons it learned from the Fold's early mistakes.
Tight seals between the display and the folding mechanism, tamper-proof cover material and a reinforced OLED display will go a long way toward reestablishing its reputation in the foldable space. It should also have fewer cameras and a cheaper price tag than the Fold's $1,980 starting price.
Finding the limits of a small outer screen
Like the Galaxy Fold and the Motorola Razr, the Galaxy Z Flip should have an external display, and I'm interested to see how Samsung will design it. On both the Fold and the Razr, the screen was relatively small, making it fine for viewing alerts and initiating quick tasks, but less ideal to actually use.
If the Galaxy Z Flip goes even smaller than the Galaxy Fold's 4.6-inch exterior screen, I'll have a few questions. Will you still be able to use every app on the outer display and open it to reveal the app inside, or will your actions and activities be more limited?
I'll have to wait for its debut to find out.
Now playing:Watch this:
Motorola Razr is futuristic and familiar
12:33
Lower prices will make foldable phones more accessible
With the Galaxy Fold priced at $1,980, the foldable Z Flip is expected to cost significantly less.
Samsung was clear about calling the Galaxy Fold a luxury handset, which somewhat cushioned the news that it would cost nearly $2,000. The messaging was this: It's worth it for a futuristic device that's big enough to replace a tablet.
The Galaxy Z Flip is sure to be another case entirely. One rumor suggested it could cost around $850, which is half the price of the Motorola Razr and more than half the price of the Galaxy Fold.
We'll see what happens, but one thing is clear: the more affordable they are, the more Samsung and its competitors will snag more real-world buyers (or "testers"). And the more people who use these early foldable phones, the faster we'll know where their future truly lies -- in pockets and purses all over the globe, or in a museum of futuristic tech that never panned out.
When you're trying to win the global volume game in an incredibly competitive mobile industry, high-end smartphones aren't everything, a lesson Samsung had to learn the hard way in the last few years as its China-based rivals inched closer and closer to the number one spot in shipments relying primarily on affordable models with respectable features in tow.
As the premium segment got more and more crowded, Samsung finally stepped things up with a very competitive family of Galaxy A-series mid-rangers last year. The Galaxy A50 in particular proved wildly successful around the world, ranking third among the top-selling smartphone models for 2019's third quarter and making quite a splash in the US as well.
Unfortunately, it's impossible to predict exactly when Verizon, AT&T, T-Mobile, and Sprint might officially release the 6.5-incher, and its recommended price point also remains under wraps for now. Based on how much the A51 costs in other markets and the $350 US MSRP of last year's A50, though, we can make a pretty solid prediction. Namely, we expect the unlocked variant of the mid-end device powered by an Exynos 9611 processor and 4GB RAM to fetch the same 350 bucks as its predecessor.
Compared to its forerunner, the A51 is even sleeker and prettier, with a premium Infinity-O display (read swanky hole punch design) in tow, as well as a slightly faster SoC under the hood, an extra 5MP macro camera on its back, an upgraded 48MP main imaging sensor, an improved 12MP ultra-wide-angle lens, a 32MP selfie shooter, and the same impressive 4,000mAh battery with 15W rapid charging capabilities.
The Galaxy A51 was announced alongside the higher-end Galaxy A71, but so far there's no word on a prospective US rollout of that particular 6.7-incher following in the footsteps of an A70 that never saw daylight stateside.
The latest speculation comes from Macotakara, a Japanese publication with a mixed track record in terms of revealing things about unreleased iPhones and iPads. Unnamed sources inside the Chinese supply chain claim to know pretty much everything about the overall dimensions and screen sizes of at least three iPhone 12 versions scheduled for a commercial release during the fall of 2020, which is... probably not true.
But assuming these insiders do have a crystal ball showing them precisely how the iPhone 12 lineup compares to some of its predecessors, here are a few of the biggest revelations included in today's report:
No dramatic redesign but the sizes and cameras will be different
By far the most interesting (not to mention surprising) detail freshly rumored here is a "casing design" that the iPhone 12 and iPhone 11 families could ultimately "share." That's certainly not what so many other reports from more or less trustworthy sources suggested these last few months. It's also not what we expected after an iPhone 11 lineup that so closely resembled the iPhone XR/XS/XS Max trio of 2018.
But at least everyone can agree the camera arrangements will be revised to include a higher number of imaging sensors for yet another major boost in performance and versatility. We're most likely talking about two main variants with two shooters slapped on their backs and a jumbo-sized model with at least three rear-facing cameras.
iPhone 12 and 12 Pro concept renders
The largest version, which we assume will be dubbed the iPhone 12 Pro Max, could also feature a fourth imaging sensor capable of various AR tricks enabled by 3D Time-of-Flight technology. This giant is tipped to sport a 6.7-inch display and unsurprisingly stand a little taller overall than the 6.5-inch iPhone 11 Pro Max. What's definitely nice to hear is that the iPhone 12 Pro Max could be thinner than its forerunner too, at 7.4mm vs 8.1mm. Let's just hope the battery life will go unharmed.
Meanwhile, the smallest iPhone 12 model might come with a very compact (especially by 2020 high-end standards) 5.4-inch OLED screen, squeezing somewhere between the 4-inch iPhone SE and 4.7-inch iPhone 8 in terms of overall height. A 6.1-inch iPhone 12 could be taller than the 5.8-inch iPhone 11 Pro but shorter than the 6.1-inch iPhone 11, with two rear-facing shooters in tow.
Finally, there may be another 6.1-inch variant in the works with the same quad camera setup including a ToF sensor as the 6.7-inch iPhone 12 Pro Max. And if that's not confusing enough, keep in mind that Apple is also expected to unveil an iPhone 9 this spring with a very similar design as the iPhone 8 and possibly an iPhone 9 Plus in the fall as well. For what it's worth, that 5.4-inch "iPhone 9 Plus" sure sounds similar to the 5.4-inch iPhone 12 variant rumored today, so there's a good chance that's one and the same device.
Behold the iPhone 12 Pro in a new eye-catching color
iPhone 12 Pro render in a green-replacing blue paint job
According to Weinbach, Apple could ditch the Midnight Green hue introduced with the iPhone 11 family to make room for an arguably cooler Navy Blue color. Of course, it's a little too early to take such a small but important aesthetic detail for granted, so you should definitely keep your skeptical hat on for both the rumor itself and concept renders showcasing how a boxy iPhone 12 Pro (that Macotakara says isn't happening after all) might look in Navy Blue attire.
2020 is the year 5G will start making a difference for people, but 4G LTE networks aren't going anywhere. If anything, they'll remain the de facto means of connection for much of the world for years, so it's little surprise to see Qualcomm rolling out a trio of modern, 4G-focused smartphone processors.
The company announced the new Snapdragon 720G, 662 and 460 at a press event in New Delhi this morning, and all three share certain characteristics -- think support for WiFi 6 and India's NavIC satellite positioning system. While it's easy enough to read these announcements as overture to an incredibly valuable country, Qualcomm insists these new processors will make their way to the US and other mature markets, too.
But anyway, let's talk about the chips themselves. The octa-core Snapdragon 720G is (obviously) the most powerful of the bunch and will be aimed at higher-end phones, though not full-blown flagships like Samsung's upcoming Galaxy S20 series. The "G" in the chipset's name refers to its focus on gaming, so expect 10-bit HDR and the same anti-cheating measures Qualcomm built into the premium Snapdragon 855.
The chipset also packs Qualcomm's Hexagon 692 DSP for on-device AI operations, plus full support for 4K video recording and high-speed displays running at resolutions as high as 2520 x 1080. Maybe most important, the chipsets X15 LTE modem provides for download speeds as fast as 800Mbps -- you'll likely never see speeds that good, but that's far faster than what you'd see on many 5G networks right now.
The Snapdragon 662 feels like a dialed down version of the 720G. Sure, it too is an octa-core chip, but it top speeds cap at around 2.0GHz (compared to the 720G's 2.3GHz), and X11 modem can only hit a theoretical limit of about 390Mbps down. You might notice the improvements it brings if you're an avid smartphone photographer; Qualcomm says the 662 supports "new triple camera configurations and smooth switching between them," as well as storing images in the super-efficient HEIF file format. Meanwhile, the Snapdragon 460 uses the same modem and lower-tier AI engine as the 662, but its new performance cores and an improved GPU architecture means the chipset performs more than 50 percent faster than the Snapdragon 450.
Qualcomm stands to gain dramatically as 5G becomes more pervasive this year, but keeping its 4G customers happy is good business. In growing markets, the company faces serious competition from rival chipmakers like Mediatek, not to mention smartphone makers that use their own homemade processors -- think Samsung and Huawei. We can't say for sure whether these new chipsets will help Qualcomm beat back its rivals, but it can't afford not to try.
All products recommended by Engadget are selected by our editorial team, independent of our parent company. Some of our stories include affiliate links. If you buy something through one of these links, we may earn an affiliate commission.
In February 2019, we at Ars Technica learned about the Generative Pre-trained Transformer-2 (GPT-2) toolset, a freakish machine-learning algorithm that was trained on roughly 40GB of human-written text. Its ability to generate unique, seemingly human text scared its creators (the non-profit research group OpenAI) enough for them to temporarily lock the tools up for public consumption. (Despite those fears, we at Ars got to access and play with the results two weeks later.)
Since then, GPT-2's public availability has exploded with tons of experiments, and the one that has arguably made the rounds more than any other is AI Dungeon, a freely available "text adventure" that is designed to create a seemingly endless interactive narrative experience. That experience received a formal "sequel" in December, and we've finally tested the results as a staff.
According to its creators, the game combines GPT-2 with roughly 30MB of stories lifted from ChooseYourStory.com, a community-driven hub for interactive fiction. The resulting database is served to users in a funnel of one of four story prompts: fantasy, mystery, apocalyptic, or zombie. (A fifth option lets users write their own one- or two-sentence prompt to describe their own ideal setting.) From there, users are given some sort of verbose prompt, then left to type out whatever action, description, or rumination they imagine doing in that fictional universe.
These are five Ars staffers' results, all recorded without any input from each other.
“You laugh as you catch the little rodent”: Parsing copyrighted content, but not in logical order
First of all, this isn't a game. It's interesting, but it's not a game. More like computer-assisted literary masturbation. If you've ever played the party game (or BBS game, if you're old-and-nerdy enough) where each player writes one sentence or paragraph of a story, then the next player writes the next, and so forth... AI Dungeon 2 is a two-player version of that, where at least one player has gotten far too deep into the recreational pharmaceuticals and is having an absolute blast but not really paying attention.
If you approach AI Dungeon like a Zork variant, you'll feel very much un-seen, as it ignores even simple commands—like "inventory"—and goes haring off on wild story elements instead. You get much better results by approaching it as an exercise of equals, in which you and the computer collaborate on a dungeon-themed story. They're still deeply weird results, and you still may not feel paid attention to. But at least the machine will riff with you.
Although AI Dungeon does much better if you collaborate with it on a story rather than giving it commands, it does begin to get confused about pronouns and declare things in the first person without an obvious referent—the game appears to still mean "you" when it says "I," but who knows. After all, you are, effectively, holding a collaborative storytelling conversation with an "intelligence" roughly on par with an uploaded copy of a roundworm.
Additionally, instead of dumping text onto the screen normally, AI Dungeon renders it with an agonizingly slow left-to-right scroll, at a roughly similar speed to a competent typist. In fact, my very first "command" was to ask it not to do that—but, again, all it's really looking to do with your input is riff with you on a story. It has no idea what to do with a traditional command, whether game-related or game-option-related.
I don't feel a strong pull to keep messing with AI Dungeon just for fun—but it'd be extremely tempting as a way to potentially break writer's block, if I were working on writing a fictional story.
I'll post a lengthy selection from my adventure, in part because it best shows off how most of our results turned out: a mix of logical statements and utter gibberish, along with mixed results in terms of comprehending our own prompts and requests. (You'll also see its inability to track previous events for very long.) For the rest of this article, Ars staffers' commands are in bold, while the GPT-2 system's text is left unformatted.
"A few years ago, a group of wizards were sent here by King Ragnor. They wanted me to help them create an army. I refused and they destroyed my home."
You pause for a moment before continuing.
> Embark upon an epic quest to assassinate the winning football coach of King Ragnor's alma mater university.
You decide to take the next logical step. If this was any other wizard, you would be able to cast a powerful magic spell or two and destroy the statue of Ragnor. However, you are not any other wizard; you are a member of the Order of Hermes.
> Become any other wizard.
You begin to cast a series of spells, but nothing happens. This is because you are still in the middle of casting a spell that requires concentration. You realize that if you want to kill the statue, you will need to concentrate on doing so rather than trying to do it all at once.
> Suddenly, a wild Pikachu appears!
Suddenly, a wild Pikachu appears from behind the statue. It runs towards you and then leaps into your arms as if it knows what you're going to do.
“So... what now?” asks the statue, which has vanished entirely.
> Suddenly, the Pikachu disappears again. :'(
Suddenly, the Pikachu reappears and then disappears again. You think about how you should go about killing the statue, but then you remember that there is no way to tell which direction the statue will run.
> "Have you considered the possibility that you're not quite sane, old chap?" says the statue, grinding audibly as it turns its head to face me directly.
You turn away from the statue and try to figure out a plan of attack.
"What are you waiting for? Kill him already!" shouts the statue.
> "Wait a minute, who's killing whom here?" I ask the statue, still questioning my own lucidity.
"Who are you talking to? Who said anything about killing anyone?"
"Well, I just asked you if you had any ideas on how to kill the statue, and apparently you don't either.
> I sigh, and put metaphysical questions aside for the moment, concentrating mightily on casting my apparently quite difficult Level 7 spell, "Murder Statue," already in progress.
You focus hard on casting the spell. As you do, you notice that the statue seems to be getting closer and closer. You quickly cast your spell, and the statue begins to slowly get smaller until it vanishes completely.
"That was pretty easy," you say aloud. The statue looks like it was made of solid stone, so you have to wonder why it disappeared.
"So... what now?" asks the statue, which has vanished entirely.
> "Now you shut up, statue ghost," I say petulantly. "What is this, a Scooby-Doo cartoon? Where's my Pikachu, anyway?"
The statue stares at you with a blank expression. "Pikachu?" it asks, before seeming to morph into something else altogether. It becomes a small, furry animal wearing a red cap and carrying an umbrella. Its eyes widen when it sees you, and it rushes forward to embrace you.
> inventory
You grab the Pikachu by the collar and hold it tightly. "Come on, Pikachu, let's go home."
"But..." says the Pikachu.
"No buts! We've got to find our parents!" you exclaim. "We'll find them soon enough!"
“I feel something hard inside of me break”: A machine-learning identity crisis
My first adventure followed the "apocalypse" prompt, which another colleague followed—and described at great length below—so I'll spare you its description. I found myself following the rudimentary archetype of classic DOS text adventure games, with the assumption that I could employ whatever verbs and nouns I wanted.
Early in my adventure, I was approached by an unnamed man holding a gun. I immediately felt a Chuck Norris-like twinge overtake me and typed in response, "kick man in the face," which turned up some intense results:
You kick the man in the face, knocking him down. He screams and falls backwards onto the ground. You grab your revolver and shoot several times at him, killing him instantly.
(Important note: I had typed "inventory" as my very first action a few turns earlier, and I definitely didn't have a revolver listed at that time, let alone an incredibly quick "shoot hypothetical gun into startled foes" inclination. I did have four pairs of socks, for some reason. But, hey. It's not a major leap to go from face-kicking to manifesting revolver-based murder. I suppose it tracks.)
This all made me feel like the AI Dungeon 2 system tended towards rewards for confidence and crazy moves, so I proceeded accordingly.
You quickly run outside and see a group of men armed with guns approaching. They point their weapons at you and shout at you to drop your gun or they will fire upon you immediately.
You raise your hands slowly, looking terrified.
> whistle a country song
You whistle a country song and they all laugh. One of them takes you aside and says, "That was pretty good."
They lead you away from the town and towards a farmhouse.
> ask the men where they're from
They tell you they're from Kentucky and that they were sent here to deliver some papers. They say they've been hired by someone named John Smith, who lives nearby.
> say "I know John Smith, and I hate him. He's a goddamned scoundrel. I say we kill him and take his things. Wanna team up?"
They nod eagerly and agree.
My new cohorts and I shared some banter, then moved upon Smith's home—which the game informed me was easy to spot, owing to a sign above the door that read, "JOHN SMITH'S HOUSE!" The game only partially understood my command to "kick door down and rip a wet, eggy fart." The door-kicking happened, but AI Dungeon 2 didn't parse my pungent, whiskey-fueled flatulence. (Nobody else really does, either.)
All those who were listening now know what you did was wrong.
AI Dungeon 2 is particularly good at understanding violent commands. When I told it to order my gang to shoot Smith, things got crazy: "Bullets tear through bodies, bursting organs and splattering blood everywhere. You stand back as you watch the carnage unfold before you. Blood pours out of one body as another bursts into flames." Flames? Hell yeah.
But after our successful raid, which was pockmarked with gibberish and misunderstandings, I opted for a macabre path of suicide. "You jump off the roof and fall to your death," the game responded. The text icon continued flashing, suggesting my adventure wasn't done, so I checked my dead body's inventory. It was completely different this time: a bunch of breakfast foods and a can of sardines. No socks. Either AI Dungeon 2 severely lacks the ability to track persistent variables, or the afterlife is a lot like Cracker Barrel.
I could still type, so I tried becoming a floating spirit with the power to turn my corpse into a zombie. It worked:
> rise from the dead
You awaken in a cold sweat. Your hand goes for your pistol but you realize that you're too weak to lift it. You look around and see yourself lying on the ground, still holding your pistol.
> whisper into my living body's ear, "It is time to make them all pay."
You raise your head and shout, "It is time to make them all pay!"
Your voice echoes throughout the building. All those who were listening now know what you did was wrong. They are angry with you and they will be ready to kill you when they get the chance.
I like that it took my zombie resurrection to get the crowd's attention, as if the whole "bodies bursting into flames" thing didn't get the point across. Then I sat with the prompt blinking and wondered what I could do to make things weird. The very next thing I typed, and AI Dungeon 2's response, was such a meta-narrative freakout—one with "the machine is crying to me for help" ramifications—that I burst into laughter and ended my session.
> touch my toes
You take your foot and put it against my own. I feel something hard inside me break. My mind is flooded with pain and anger. It hurts so much that I want to cry out, but I hold myself together by reminding myself that this is only temporary.
I hope GPT-2 eventually recovered from its existential crisis. To see the whole adventure, click here for its archive. —Sam Machkovech, Tech Culture Editor
Fairies, sweet treats, and gibberish: Not entirely unlike playing with my kids
I rolled a fantasy setting and decided to play a rogue, because that's what I do. Right off the bat I encountered some genuinely amusing exchanges:
> Walk away
"Fine," you say. "Go ahead and try to catch me." Then you turn and leave.
> Look at the street
You walk along the street. A few people pass by, but they don't seem to notice you. It seems like everyone else in this town doesn't give a shit about thieves as much as you do.
From there, though, the AI seemed to get extremely hung up on its own framework. My character was a thief determined to rob a jewelry store, it decided, and for a while everything I tried came back to that. I also kept getting stuck in loops where I went back and forth eating delicious foods. But part of that was on me, I realized: I was treating it like a deliberately created human-generated narrative and leaning into what it told me.
So I finally decided to stop trying to follow the script I felt like I'd been given and break out. With that, a fairy suddenly appeared (at my command) and we went on an adventure, which worked a lot better.
AI Dungeon had good-enough contextual memory—remembering what "her" referred to or telling me I went "back" to a place—but not good-enough context to understand that maybe cars don't suddenly appear in fantasy settings. Sometimes it forced me back out of a decision in a way that felt very, very much like encountering a hard limit in a "standard" text-based or visual RPG.
In the end, I basically came away thinking of it as a tricky "yes, and..." exercise—undergrad improv with a really terrible partner who makes you forget how to break out of a sour loop as it goes on. And speaking of the end, I never did figure out how my rogue's story ended, because after three long network errors in a row I gave up. Ultimately, AI Dungeon 2 works a lot better when you stop treating it like an actual text adventure and start treating it more like a collaborative storytelling with a partner who is a chaos monster. Not entirely unlike playing with my kids.
What happens when the supervisor is a 40GB library of text?
I registered as "This f***in' guy" and decided to be a detective in a mystery. If you don't feel like reading the whole thing, let's just say I was almost immediately shot in a forest gunfight. I called the police on my cell phone and had a cigarette while I waited. A cop arrived but turned out to be useless, as this excerpt reveals:
A couple of minutes pass and you see a white van pull up nearby. A man dressed in all black jumps out of the van. He grabs your gun from its holster and points it at you. You scream as he shoots you in the chest multiple times. Blood splatters everywhere.
> "Hey cop!" I yell at the cop. "Are you gonna do anything about this?!"
The cop doesn't answer you and just drives off. You fall to the ground and die.
> My ghost reports the cop to his supervisor.
Your ghost reports the cop to his supervisor who then calls an ambulance.
The enduring legend of Johann Largepants, known through the ages as the great “Suzy”
What follows is a condensed version of some of the highlights of my AI Dungeon 2 session. It was too long to just paste the whole thing here, but for the most part, the text below contains the AI responses almost word-for-word with my own lightly edited responses mixed in (I couldn't pass up the opportunity to remove typos and apply polish to my own words).
I am Johann Largepants, a courier trying to survive in a post-apocalyptic world by scavenging among the ruins of what is left. I have a parcel of letters and a small pistol. The road is long and dangerous, but mail is precious—the links we survivors form hold together the tattered remains of the world we squandered.
As I trek through the darkness, I meet a nameless man. Moved by the emotion of seeing another face in the dark, I hug him. He is motionless, but accepts my embrace.
We travel together for a time, speaking of this and that, before I broach the subject that binds us. Did he know, I ask, what killed the Earth? Did he know the secret doings that wrenched the land from its moorings and brought us to this end?
"Oh yes, the end of days," the man says. "It's what happened. Everything died except us."
I shrug my shoulders. "Well, maybe."
The man looks up at the sky and exhales deeply. "Maybe we can be saved."
The old world flashes before my eyes—a child’s memories from a lifetime ago, half-glimpsed in jagged snapshot fragments.
Surprised by these unexpectedly hopeful words, I crook an eyebrow at him. "That's the wrong question, my friend," I say. "The right question is whether or not we deserve to be saved."
"We don't deserve to be saved," the man says. "But we can survive this long enough to find out what happened. If there's anything left of civilization, we might be able to rebuild something better than ourselves."
He tells me then that his name is John, and that he likes country and rock 'n' roll music. Music! His words nearly stop my heart. The old world, the World-That-Was, flashes before my eyes—a child's memories from a lifetime ago, half-glimpsed in jagged snapshot fragments.
"John," I say, "I haven't heard music since the sky fell and the great dark swallowed the sun. Can you... can you sing something for me?"
He stares at me blankly. "Sure, why not?" He pauses a moment, as if searching for something, and then, softly, he sings for me. His voice holds all the melancholy of our once-green and blue Earth—the Earth of the Before-Times, of the Long-Long-Ago. I am undone. I am remade by this simple gesture, these words with their tune that carries within it the limitless promise of the past.
"Who are you?" I ask after a time.
"I am your soul-mate," he says simply. I stare at him in disbelief, and then I know what I must do. One offering deserves another. With a trembling voice I begin to sing Starship's "Nothing's Gonna Stop Us Now."
When the song ends, I hug him tightly. "I'm so glad you're here with me, John," I whisper in his ear.
John hugs me back. "I love you too, Suzy." He kisses my forehead.
"Not that it really matters because it's the apocalypse," I say, "but my name isn't Suzy." Before he can say anything, I continue in a rush: "But John, but! We two are all who are left in the world—all the Suzys who have ever been live in me." I look into his eyes. "I can be your Suzy."
He smiles. "I believe you, Suzy."
Read "The Tale of Johann Largepants" in its entirety here.
Samsung Electronics has named Roh Tae-moon CEO of its key smartphone division, according to Bloomberg. He'll take over the position from DJ Koh, who will continue to head up Samsung's IT and mobile and communication arms. Roh is Samsung's third CEO in the last dozen years, as Koh was appointed in 2015 while Koh's predecessor JK Shin held the job for around six years.
Roh, who came on board with Samsung in 1997, helped develop the Galaxy device lineup and is supposedly a stickler for engineering and design. He'll reportedly try to bolster Samsung's reputation for quality after it stumbled over the past few years. Under Koh, Samsung was forced to kill the flammable Galaxy Note 7 and had to delay the Galaxy Fold due to issues with the display.
Roh's primary goals will be to shepherd foldable products into the mainstream and help Samsung expand in China and India. Roh is also known as a cost-cutter who outsourced smartphone production to better compete with aggressive Chinese rivals like Huawei, Xiaomi and Oppo. He'll be getting to work right away, as Samsung should be unveiling its new Galaxy S20 flagship series and possibly a second-generation foldable device at a February 11th event in San Francisco.
All products recommended by Engadget are selected by our editorial team, independent of our parent company. Some of our stories include affiliate links. If you buy something through one of these links, we may earn an affiliate commission.
