Emilia David | The Verge

AI design apps made my new apartment look odd

2024-07-11T08:30:00-04:00

When I moved to a new studio apartment last year, it was my chance to live out my DIY YouTube girly dreams and design it to my heart’s content. But it turned out to be harder than I thought. Since I couldn’t afford an actual designer, I decided to try out some of the generative AI-powered design apps I’d seen floating around the internet.

AI-based design tools began cropping up about the time ChatGPT burst onto the scene. They come in different flavors, from platforms where you upload a photo and write a prompt for the AI to overlay a new image on top of to ones that suggest new styles for you to try.

One of the images I asked AI apps to redesign. | Photo by Emilia David / The Verge" data-portal-copyright="Photo by Emilia David / The Verge" />The second photo I uploaded, showing a corner of my living room. | Photo by Emilia David / The Verge" data-portal-copyright="Photo by Emilia David / The Verge" />

I decided to try out a couple of straight AI chatbots (ChatGPT and Gemini), a retail-based AI assistant from Ikea, and three design apps (Spacely AI, Decoratly, and RoomGPT). I uploaded photos of my apartment to the platforms and wrote two prompts for those that had a prompt box: “Give me a storage solution for this area” and “Transform this image into a midcentury modern-style living room.”

Here are short summaries of how each one fared.

ChatGPT and Gemini

My conversation with ChatGPT about redesigning my apartment. | Screenshot: ChatGPT" data-portal-copyright="Screenshot: ChatGPT" />

ChatGPT and Gemini were (obviously) not made explicitly for design, so the most I expected were some suggestions and maybe an edit of the photo with some information about the items it chose.

I got some of what I hoped for. Both ChatGPT and Gemini gave me storage suggestions, with ChatGPT telling me what materials I should look for to keep the room in the midcentury modern style. Neither chatbot was capable of changing my photo or generating its own living room designs in my chosen style.

ChatGPT is free to use for a limited number of messages; otherwise, it’s $20 per month. Gemini is free, but the advanced version with improved AI models is $19.99 with a Google One membership.

The Ikea AI Assistant

Ikea’s chatbot gave me some furniture ideas. | Screenshot: Ikea" data-portal-copyright="Screenshot: Ikea" />

Ikea created a custom version of ChatGPT last February so that shoppers could ask questions about furnishing their living spaces and get suggestions about styles and furniture. I uploaded a photo of a corner of my living room, which admittedly showed a messy pile of workout stuff, vinyl records, a bookshelf, and just general bric-a-brac, and checked out its suggestions.

To store my yoga mat (and a travel pillow it mistook for a yoga mat), Ikea’s chatbot suggested I get a storage rack and other “decorative items.” (It also suggested that I add a bookshelf even though there was already one in my photo.)

As expected, after suggesting storage solutions, Ikea wanted me to buy its products, so I gave it rough measurements of the spot and told it I would love items evoking a midcentury modern feel but with dark wood. It responded with photos of the items and told me where to find them. Despite all that, it still felt more like a search tool than a design app.

The better option is probably Ikea’s non-ChatGPT-based mobile app, which uses augmented reality to help you imagine what your space would look like by overlaying its product in your home.

Ikea’s custom GPT is free at the OpenGPT Store.

Spacely AI

Spacely did an okay job, but for some reason, it turned my lemon into a blue egg. | Screenshot: Spacely AI" data-portal-copyright="Screenshot: Spacely AI" />

One of the most recommended AI-based design platforms on social media is Spacely AI. After uploading a photo or choosing from a template, users can redesign a space, furnish an empty room, or edit a photo through written prompts.

I asked Spacely to reimagine my space in a midcentury modern design with mainly wooden furniture. Spacely is more customizable than other platforms, allowing me to control how much its model follows my prompts (such as preferred style, color palette, etc.). However, the customization choices are vastly limited in the free version; if you want to do more than just try it out, you’ll need to go for a paid plan.

Spacely had a basic understanding of what I wanted, but the images it generated didn’t really fulfill my brief. For example, I happened to upload a photo that included two plastic containers and a lemon, and the AI generator transformed the two objects into… decorative items, I guess. The plastic containers became wooden cylinders and the lemon either is a rock or rotten fruit. (Unfortunately, it’s still normal for funny things to show up in AI-generated images.)

Spacely AI Pro is $20.75 a month for a yearly plan or $39 for a monthly subscription for unlimited prompts, watermark-free photos, and high-resolution downloads.

Decoratly

Decoratly’s stab at designing my apartment felt closer to a real room with a distinct style. | Screenshot: Decoratly" data-portal-copyright="Screenshot: Decoratly" />

Decoratly also transforms photos into a specific style. It’s very limited for free users; before I subscribed, I could only upload my photo and tap the quick redesign button to generate a generic design filled with white and black furniture and zero character.

When I upgraded to a Pro account, I was able to use Decoratly’s “Build a Prompt” feature and its image filter, which lets you give instructions on what you’d like the app to make. Unlike the other AI prompt builders I tried, Decoratly wouldn’t let me write my own prompts. Instead, I had to choose from a prepared set of words to describe what style, color, material, and texture I wanted to see in the transformed photo.

I chose the words “midcentury modern,” “dark,” “gray,” “wooden,” “metal,” “smooth,” and “neutral” for my room. The new photo it created felt closer to a real room with a distinct style than those I got with the other apps, although some of its choices could be weird — like putting some sort of table thing on top of the cylinder it transformed my electric fan into. Also, it put my monstera plant in a teeny-tiny pot that would have fallen over in five seconds flat.

Decoratly costs $12 / month for unlimited designs and additional features. A 24-hour ($3) and seven-day ($6) trial is available.

RoomGPT

RoomGPT’s attempt added a blur filter to my apartment. | Screenshot: RoomGPT" data-portal-copyright="Screenshot: RoomGPT" />

Out of the dedicated AI design platforms I tried, RoomGPT was the most disappointing.

I felt like the app did the bare minimum in redesigning my space. It did change a couple of items to fit the brief — for example, adding a couch to a room that didn’t have one — but it also removed my TV and media console entirely and never transformed the room to fit the style I wanted.

RoomGPT runs on a credit system, where each render is one credit. The free version offers two free credits. After that, there are three paid tiers based on the number of credits or room designs: $9 for 30 room designs; $19 for 100 designs; and $29 for 200 credits.

Waiting for better

In short, none of the AI apps I tried really helped me design my place. The most they did was show me the kinds of furniture that could fit the vibe I was looking for, which I could’ve done with a quick Google search anyway. None of them were capable of figuring out a new style for my space or truly reimagining my apartment. As with other things, AI is not really ready to design our living spaces.

Where the CHIPS Act money has gone

2024-06-07T08:00:00-04:00

Since the passage of the CHIPS and Science Act last August, eight companies have already received more than half of the planned government direct funding.

These companies have collectively received $29.34 billion in funding through the CHIPS Act for semiconductor factories across the country. The law, a $280 billion package to support innovation in the US, includes $52 billion in subsidies for semiconductor manufacturing and was passed last year.

These investments only concern the construction or expansion of semiconductor fabrication facilities and do not include government funding for other chip research facilities.

As of writing, Intel, Micron, Global Foundries, Polar Semiconductor, Taiwan Semiconductor Manufacturing Company (TSMC), Samsung, BAE Systems, and Microchip Technology have been the direct beneficiaries of the law.

These include projects like Intel’s factories in Arizona, New Mexico, and Oregon, the $20 billion fab in Ohio, and Micron’s $ 100 billion plant in Syracuse, New York, to build memory chips.

Intel received the biggest direct investment through the CHIPS Act, with $8.5 billion for its semiconductor projects. TSMC received $6.6 billion in funding, while Samsung rounded out the top three with $6.4 billion from the US government.

The CHIPS Act is intended to restart the US semiconductor industry and start competing against Chinese dominance in the chip manufacturing space. However, the amount set aside to jumpstart the industry cannot be the only source of capital to bring the US up to speed, according to Commerce Secretary Gina Raimondo, as the goal of the law “was never to provide the semiconductor industry with every dollar it requests.”

Raimondo has said leading-edge chip manufacturers have asked for $70 billion in funding for chip fabrication, more than the government initially expected to spend. She said the department is prioritizing projects that will be operational by 2030 and some “very strong” proposals from companies may never get funding through the act.

Fabricating chips is an expensive affair. TSMC, one of the CHIPS Act beneficiaries, earmarked $44 billion in 2022, up from $31 billion in 2021, just to expand its chip-making capacity.

The Semiconductor Industry Association says in an email to The Verge that the industry garnered more than $450 billion in private investments after the announcement of the CHIPS Act, and it expects it to grow even further.

Demand for chips has grown as generative AI models, which are primarily trained using powerful chips, have also grown in prominence. The US wants to start providing more high-powered chips and even start making next-generation semiconductors. The Biden administration announced in February that it will also start funding research into substrate packaging technologies, which would help create more leading-edge semiconductors.

FTC and DOJ reportedly opening antitrust investigations into Microsoft, OpenAI, and Nvidia

2024-06-06T12:02:34-04:00

The Federal Trade Commission (FTC) and the Department of Justice (DOJ) reportedly agreed to split duties in investigating potential antitrust violations of Microsoft, OpenAI, and Nvidia, The New York Times reported.

The DOJ will lead inquiries into Nvidia, while the FTC will look into the deal between OpenAI and its largest investor, Microsoft.

The FTC began looking into potential antitrust issues related to the investments made by technology companies into smaller AI companies in January this year. The agency sent letters to Alphabet, Amazon, Anthropic, Microsoft, and OpenAI. Alphabet, the parent company of Google, and Amazon are both investors in Anthropic.

The FTC is already investigating OpenAI’s data collection practices, having opened an inquiry in 2023 to determine if the company has caused potential harm and spread false information about individuals. Microsoft’s investment into OpenAI may also face scrutiny outside of the US, as the European Commission and the UK’s Competition and Markets Authority are separately looking into Microsoft’s $13 billion investment in the ChatGPT maker.

Nvidia, however, had not figured into any antitrust conversations in the US. Nvidia is considered the dominant figure in making the chips that power much of the AI boom. Its H100 GPUs are highly sought after, sending the company’s valuation soaring. French antitrust authorities raided Nvidia’s offices in France in connection with an anticompetitive practices investigation in September 2023.

The Verge reached out for comments from Microsoft and Nvidia but have not heard back. The FTC and DoJ declined to comment. OpenAI also declined to comment.

An investigation by the two agencies does not mean the Biden administration is opening cases against the three companies, but the Times pointed out that a similar deal in 2019 led to the government’s cases against Google, Apple, Amazon, and Meta.

Update, June 6th: Added that the DoJ and OpenAI declined to comment.

Google makes its note-taking AI NotebookLM more useful

2024-06-06T07:00:00-04:00

Google launched its note-taking app NotebookLM last year for researchers, students, and anyone who needs to organize the information they’ve gathered. Now, users can now upload Google Slides and web URLs as sources, not just the Google Docs, PDFs, and text files it accepted previously.

The new Notebook Guide also reads sources in NotebookLM and creates study guides, FAQs, or briefing documents, and inline citations can point to your own sources to fact-check AI responses — up to 50 sources per “notebook,” or project, and each source can be 500,000 words long. Previously, people could only upload five sources.

Users can also now ask questions about charts, images, and diagrams they uploaded to the platform because NotebookLM is running on Google’s Gemini 1.5 Pro, the latest large language model that currently powers the paid version of the Gemini chatbot. I was able to try my hand at these features to see how they work.

Google’s NotebookLM adds new features like a Notebook Guide that distills information into an FAQ, study guide, or briefing report. | Image: Google" data-portal-copyright="Image: Google" />

In a briefing, Raiza Martin, senior product manager at Google Labs, told reporters that NotebookLM “is a closed system.” It will not do any web searches beyond reading the website content users add. Martin says NotebookLM’s answers to queries about data or images will only come from the user’s “corpus” or body of information they add to the platform.

I tried NotebookLM to see the new features in action. The Notebook Guide was not yet available for reporters to try, but I was able to add new data sources, get inline citations, and get Gemini 1.5 Pro to look at graphs for me. I asked NotebookLM to give me information from a PDF of a line graph, and it gave me the numbers I was looking for. I also asked it to summarize the text of the EU AI Act, and it was able to give me an overview and include citations so I knew where it was pulling its answers from.

Unfortunately, web URL sources didn’t work in my demo: whenever I pasted a link into NotebookLM, the model would start to upload the website, but it didn’t show up on my list of sources.

NotebookLM is not a tool that will write research papers for you, unlike Perplexity’s Pages, which supposedly helps researchers find data and make it easier to share information (but, in my opinion, fails to do so).

Google gave examples of how people have been using NotebookLM, including shouting out author Walter Isaacson, who the company says used the platform to analyze Marie Curie’s journals for his next book. Google also says that nonprofits use NotebookLM “to identify needs in underserved communities and organize information for grant proposals.”

Martin says that while NotebookLM’s target audience remains researchers, students, and often writers, the company found other use cases, like a Dungeons & Dragons dungeon master who used NotebookLM to prepare a campaign.

NotebookLM is now available in over 200 countries and territories and supports more than 100 languages.

Asana’s new ‘AI teammate’ can tell people what to do at work

2024-06-05T09:00:00-04:00

Asana’s platform already helps teams lay out tasks and see who is assigned which responsibilities, and now it says “AI teammates” will step in with advice and plans for who should work together to get things done.

In a press release, Asana says its AI model can use stored information about the historical relationships and past projects of teams to assign work to the people with the best-matched skill sets, like tagging designers who know brand styles to work on creative projects. It can also look for missing information before responding to a request and proactively reach out to gather that data for the team.

According to Asana co-founder and CEO Dustin Moskovitz, “We’re able to do this better than anyone else because we built Asana on the Work Graph, which maps the relationship between the work your team does, the information about that work, and the people doing the work.”

Asana’s release didn’t go into specific examples but said customers are already testing it, like one unnamed marketing organization where it says the AI bot is putting together tailored marketing content, translating assets into different languages, and standardizing workflows.

The company also announced a chat interface for the AI teammate so users can ask it questions about their current project.

Asana’s AI can answer employee questions on their company. | Screenshot: Asana" data-portal-copyright="Screenshot: Asana" />

AI teammate feature is the latest entry in the AI assistant for your job genre of generative AI use cases. Google announced an AI Teammate for Google Workspace in May that users can customize to automate some work tasks. Amazon’s Q learns company data so workers can just ask it questions about their business, and Microsoft’s Copilot AI agents automate many workflow jobs without the need for a prompt.

Former OpenAI employees say whistleblower protection on AI safety is not enough

2024-06-04T12:53:59-04:00

Several former OpenAI employees warned in an open letter that advanced AI companies like OpenAI stifle criticism and oversight, especially as concerns over AI safety have increased in the past few months.

The open letter, signed by 13 former OpenAI employees (six of whom chose to remain anonymous) and endorsed by “Godfather of AI” Geoffrey Hinton, formerly of Google, says that in the absence of any effective government oversight, AI companies should commit to open criticism principles. These principles include avoiding the creation and enforcement of non-disparagement clauses, facilitating a “verifiably” anonymous process to report issues, allowing current and former employees to raise concerns to the public, and not retaliating against whistleblowers.

The letter says that while they believe in AI’s potential to benefit society, they also see risks, such as the entrenchment of inequalities, manipulation and misinformation, and the possibility of human extinction. While there are important concerns about a machine that could take over the planet, today’s generative AI has more down-to-earth problems, such as copyright violations, the inadvertent sharing of problematic and illegal images, and concerns it can mimic peoples’ likenesses and mislead the public.

The letter’s signees claim current whistleblower protections “are insufficient” because they focus on illegal activity rather than concerns that, they say, are mostly unregulated. The Department of Labor states workers reporting violations of wages, discrimination, safety, fraud, and withholding of time off are protected by whistleblower protection laws, which means employers cannot fire, lay off, reduce hours, or demote whistleblowers. “Some of us reasonably fear various forms of retaliation, given the history of such cases across the industry. We are not the first to encounter or speak about these issues,” the letter reads.

AI companies, particularly OpenAI, have been criticized for inadequate safety oversight. Google defended its use of AI Overviews in Search even after people claimed it was giving people dangerous, though hilarious, results. Microsoft was also under fire for its Copilot Designer, which was generating “sexualized images of women in violent tableaus.”

Recently, several OpenAI researchers resigned after the company disbanded its “Superalignment” team, which focused on addressing AI’s long-term risks, and the departure of co-founder Ilya Sutskever, who had been championing safety in the company. One former researcher, Jan Leike, said that “safety culture and processes have taken a backseat to shiny products” at OpenAI.

OpenAI does have a new safety team, one that is led by CEO Sam Altman.

Even the Raspberry Pi is getting in on AI

2024-06-04T03:00:00-04:00

As the AI craze continues, even the microcomputer company Raspberry Pi plans to sell an AI chip. It’s integrated with Raspberry Pi’s camera software and can run AI-based applications like chatbots natively on the tiny computer.

Raspberry Pi partnered with chipmaker Hailo for its AI Kit, which is an add-on for its Raspberry Pi 5 microcomputer that will run Hailo’s Hailo-8L M.2 accelerator. The kits will be available “soon from the worldwide network of Raspberry Pi-approved resellers” for $70.

Hailo CEO and co-founder Orr Danon tells The Verge that its accelerator’s “power consumption is below 2W and is passively cooled.” The accelerator offers 13 tera operations per second (TOPS), which is lower than chips planned for AI laptops like Intel’s 40 TOPS Lunar Lake processors.

Most AI applications run on the cloud because they often require massive amounts of energy and computing power to work. However, there has been a move to make smaller AI models and processors that require less power to bring AI to portable devices. That way, laptops and phones can run coding assistants or AI-powered photo editing applications without needing to do an API call.

AI on PCs and other portable devices is the big trend right now as many hardware makers seek to cash in on the AI demand. Microsoft revealed its laptop partners will release Copilot Plus PCs, which have built-in AI features like the controversial Recall. AMD announced new AI-branded Ryzen processors to point out that its flagship chips can now run generative AI workloads. Nvidia, which makes the much sought-after H100 GPUs that train large language models like GPT-4o, will also ship AI chips inside laptops.

Amazon’s Project PI AI looks for product defects before they ship

2024-06-03T14:21:31-04:00

Amazon’s Project PI, or “Private Investigator,” setup combines generative AI and computer vision to “see” damage on products or determine if they are the wrong color or size before the item gets sent to customers.

The way it works is that products on their way to customers go through a tunnel that scans the items. The computer vision program — a type of AI that looks at images and understands what’s in them — checks to see if there is damage. If it finds something, that item is isolated, while the system evaluates the defect and determines if there’s an issue with similar items to track down the root cause.

An example of how Project PI works. | Image: Amazon" data-portal-copyright="Image: Amazon" />

According to Amazon, Project PI is active in “several” North American warehouses and will be added to more sites throughout the year. Last year, Amazon rolled out a different system that flags frequently returned items to highlight items that tend to have issues before customers actually hit the order button. This all highlights how avoiding a potentially “nightmarish” return process is good for customers, Amazon, and the environment in terms of reducing carbon emissions.

The company says human Amazon employees review the things Project PI flagged so they can decide if they will be sold at a discounted price on Amazon’s Second Chance resell site or be donated elsewhere.

Amazon is also working on bringing in a multimodal large language model to investigate why customers are dissatisfied with the items they receive. The AI tool reviews what customers say in their feedback and then scans images from Project PI and other data sources to try to find out where things went wrong. Amazon says this technology could be helpful to its other sellers so they know if they accidentally mislabeled items.

ElevenLabs’ AI generator makes explosions or other sound effects with just a prompt

2024-05-31T15:14:00-04:00

ElevenLabs already offers AI-generated versions of human voices and music. Now, it will let people create sound effects for podcasts, movies, or games, too. The new Sound Effects tool can generate up to 22 seconds of sounds based on user prompts that can be combined with the company’s voice and music platform, and it gives users at least four downloadable audio clip options.

The company says it worked with the stock media platform Shutterstock to build a library and train its model on its audio clips. Shutterstock has licensed its content libraries to many AI companies, including OpenAI, Meta, and Google.

pic.twitter.com/pxrJy3BPOB
— ElevenLabs (@elevenlabsio) May 31, 2024

Sound Effects is free to use, but paid tiers can use the generated audio clips with commercial licenses, while free users “must attribute ElevenLabs by including ‘elevenlabs.io’ in the title.” ElevenLabs users have a set character count limit when writing prompts, with free users getting 10,000 characters per month. For Sound Effects, ElevenLabs says on its FAQs page that it will take 40 characters per second from the allotment if users set the audio clip duration themselves. If using the default audio duration, each prompt request will be charged 200 characters.

Libraries with sound effect clips already exist in the market for creators, filmmakers, and video game developers. But sometimes, these can be expensive or have trouble surfacing just the right type of sound. ElevenLabs says in its blog post that it designed Sound Effects “to generate rich and immersive soundscapes quickly, affordably and at scale.”

Other AI developers are also developing their own text-to-sound generators. Stability AI released Stable Audio last year, which creates audio clips of music and sound effects, and Meta’s AudioCraft models generate natural sound (think background noises like wind or traffic).

Perplexity will research and write reports

2024-05-30T16:55:50-04:00

AI search platform Perplexity is launching a new feature called Pages that will generate a customizable webpage based on user prompts. The new feature feels like a one-stop shop for making a school report since Perplexity does the research and writing for you.

Pages taps Perplexity’s AI search models to find information and then creates what I can loosely call a research presentation that can be published and shared with others. In a blog post, Perplexity says it designed Pages to help educators, researchers, and “hobbyists” share their knowledge.

Users type out what their report is about or what they want to know in the prompt box. They can gear the writing more toward beginners, expert readers, or a more general audience. Perplexity searches for information, then begins writing the page by breaking down the information into sections, citing some sources, and then adding visuals. Users can make the page as detailed or concise as they want, and they can also change the images Perplexity uses. However, you can’t edit the text it generates; you have to write another prompt to fix any mistakes.

I tried out Pages ahead of time to see how it works. Pages is not geared toward people like me who already have an avenue to share our knowledge. But it doesn’t seem geared toward researchers or teachers, either. I wanted to see how it can break down complex topics and if it can help with the difficult task of presenting dense information to different audiences.

Among other topics, I asked Perplexity’s Pages to generate a page on the “convergence of quantum computing and artificial intelligence and its impact on society” across the three audience types. The main difference between audiences seems to be the jargon in the written text and the kind of website it takes data from. Each generated report pulls from different sources, including introductory blog posts like this one from IBM. It also cited Wikipedia, which drove the student report vibe home.

One of the pages Perplexity generated for me. | Screenshot: The Verge" data-portal-copyright="Screenshot: The Verge" />

The Perplexity-generated page did a passable job of explaining the basics of quantum computing and how AI fits into the technology. But the “research” didn’t go as deep as I could have if I were writing the presentation myself. The more advanced version didn’t even really talk about “the convergence of quantum computing and AI.” It found blog posts talking about quantum inflection points, which is when quantum technologies become more commercially viable and is not at all related to what I asked it to write about.

Then, I asked Pages to write a report about myself, mainly because the information there is easily verifiable. But it only took information from my personal website and an article about me on my high school’s website — not from other public, easily accessible sources like my author page on The Verge. It also sometimes elaborated on things that had nothing to do with me. For example, I began my journalism career during the 2008 financial crisis. Instead of talking about the pieces I wrote about mass layoffs, Perplexity explained the beginnings of the financial crisis.

Pages does the surface-level googling and writing for you, but it isn’t research. Perplexity claims that Pages will help educators develop “comprehensive” study guides for students and researchers to create detailed reports on their findings. I could not upload a research paper for it to summarize, and I couldn’t edit the text it generated, two things I believe users who want to make the most of Pages would appreciate.

I do see one potential user for Pages, and it isn’t one Perplexity called out: students rushing to put out an assignment. Pages may improve in the future. Right now, it’s a way to get easy, possibly correct surface-level information into a presentation that doesn’t really teach anything.

Pages will be available to all Perplexity users, and the company says it’s slowly rolling it out to its free, Pro, and Enterprise users.