Hand Pressing Buttons (CSA Images/Getty Images)

Threads’ chaotic “For You” feed quantified

How old are the posts Threads are recommending to its users?

8/27/24 12:20PM

Since Twitter's rapid unscheduled disassembly following Elon Musk's purchase of the platform, hundreds of millions of users have turned to Meta’s Threads as a potential replacement for the newsy town square experience, minus the hate speech and crypto scams.

The year-old platform now has over 200 million users and has been rapidly adding features to achieve golden age Twitter product parity. While users have been praising the improved quality of discourse, and a seemingly functional content moderation team, one the most frequent complaints has been how bad the default “For You” feed view is for following breaking news, something that Twitter (in the pre-Musk days) excelled at.

Threads users routinely mock what seems like a pretty consistent 48 hour delay for event-specific posts, which is especially frustrating as we live through one the most chaotic election years in recent memory.

Post from @miasato.2 — (Threads post from @miasato.2)

To try and better understand how much the algorithmic “For You” feed varies from the more straightforward, reverse-chronological “Following” feed, I analyzed 300 posts from both feeds from my Threads account.

No recent posts For You

When I plotted out the age of each post from my “For You” 300 post sample, indeed it does show a chaotic scramble.

If there was an earthquake in your area, and you jumped onto Threads on your phone, the default view you see is “For You,” which includes accounts you don’t follow. Based on my sample, it’s very unlikely any urgent posts on breaking news would show up until hours later.

Over one-third of the posts in my “For You” feed were between six and twelve hours old. Only 12% of the posts happened within an hour of seeing them in my feed. If you really want to follow what the chatter is surrounding a breaking news event, you should be looking at your “Following” feed. Of course, that relies on you having a well-curated list of people who follow and share news.

When I plotted the posts from my “Following” feed, which includes posts from the 954 accounts that I follow, you can see a very different pattern. A clear line of posts going up and to the right, indicating a reverse-chronological series of posts getting older as I scrolled down my feed. 38% of the posts in this feed were posted within an hour, and the vast majority were less than six hours old. Much more useful for breaking news.

Stubborn default settings

Users have been clamoring for a way to control what their default feed view is, but Meta loves the algorithmic feed, so in order to swap views, you need to know where to find the controls for this, which isn't exactly obvious. On mobile, you have to reveal the hidden controls by tapping the Threads logo. It's a bit easier on the web, where you just click on the drop down menu at the top of the feed. On the web, you can also just bookmark https://www.threads.net/following to get straight to the most recent stuff.

Another thing that makes the stubborn initial “For You” feed problematic, is that by default, Threads limits recommending any “political content” from users that you aren't following. You have to dig into your settings to opt-in to see recommended political posts from accounts you don't follow. This decision has been criticized by political observers and creators, especially during this intense election year. The company does not limit political posts in your “Following” feed.

This dataset is admittedly a small sample, of just one account at one moment in time, so your mileage may vary. Meta says that Threads' AI system “blends” content from recommended posts and accounts you follow, and considers your inferred interests and your behavior on the platform to decide what to show you.

A quilt of Threads posts. — Some of the 300 Threads posts in this analysis. (Image: Jon Keegan)

But you will have to take Meta's word on it. Earlier this month Meta shuttered its CrowdTangle tool, which provided researchers with a crucial view into what content is shared on Facebook and Instagram. Meta recently released a Threads application programming interface, but right now it mainly enables automated posting and a platform-wide analysis isn't yet possible.

Like the rest of Meta's algorithmic systems, it continues to be a black box.

Meta did not respond to a request for comment.

Rani Molla17m

Elon Musk gives an estimate for Tesla’s AI6 chip timeline... while the AI5 is still unfinished

Tesla CEO Elon Musk said yesterday that the company’s AI6 chip could, with “some luck and acceleration using AI,” be finalized and sent to manufacturing by December. For those paying attention, Tesla hasn’t confirmed that its previous chip, the AI5, has reached tape-out, with Musk saying only that the design is in “good shape” and “almost done.” Still, Musk is already talking about subsequent chips AI6, AI7, AI8, and beyond.

Here’s a roundup of when these chips are expected, what they’re supposed to do, and what Musk himself has said about them.

While the AI5 and AI6 will be made by TSMC and Samsung, respectively, Musk has said Tesla eventually aims to manufacture its future AI chips at Tesla’s upcoming Terafab factory in Austin.

Rani Molla3h

NHTSA expands Tesla FSD probe, focusing on whether system can detect when cameras can’t see the road

The National Highway Traffic Safety Administration said it is expanding its probe into Tesla’s Full Self-Driving system into an engineering analysis covering about 3.2 million Teslas, a majority of its vehicles that are on the road in the US, Reuters reports.

The agency is focusing on Tesla’s “degradation detection system,” which is meant to recognize when its camera-based technology cannot reliably perceive the road and prompt drivers to intervene:

“Available incident data raise concerns that Tesla’s degradation detection system, both as originally deployed and later updated, fails to detect and/or warn the driver appropriately under degraded visibility conditions such as glare and airborne obscurants. In the crashes that ODI has reviewed, the system did not detect common roadway conditions that impaired camera visibility and/or provide alerts when camera performance had deteriorated until immediately before the crash occurred.”

Tesla CEO Elon Musk has long argued that the company’s self-driving approach does not require the expensive lidar sensors used by rivals such as Waymo.

NHTSA deepens probe into Tesla's driver-assistance system

The agency is focusing on Tesla’s “degradation detection system,” which is meant to recognize when its camera-based technology cannot reliably perceive the road and prompt drivers to intervene:

“Available incident data raise concerns that Tesla’s degradation detection system, both as originally deployed and later updated, fails to detect and/or warn the driver appropriately under degraded visibility conditions such as glare and airborne obscurants. In the crashes that ODI has reviewed, the system did not detect common roadway conditions that impaired camera visibility and/or provide alerts when camera performance had deteriorated until immediately before the crash occurred.”

Tesla CEO Elon Musk has long argued that the company’s self-driving approach does not require the expensive lidar sensors used by rivals such as Waymo.

$1B

Rani Molla3h

Apple is behind the rest of Big Tech when it comes to developing its own AI, but that hasn’t stopped it from cashing in on the AI boom. The iPhone maker stands to bring in more than $1 billion in App Store fees this year from other companies’ generative-AI apps, mostly from ChatGPT, The Wall Street Journal reports, citing data from App Magic.

Unlike rivals pouring hundreds of billions into AI infrastructure, Apple’s spending has been relatively modest, with its overall capital expenditure actually declining last quarter. Its lucrative App Store model lets Apple profit from AI as a gatekeeper without fully joining the expensive race to build it.

OpenAI is shipping everything. Anthropic is perfecting one thing.

The two AI titans are in a race to grow revenues, but they have very different strategies for releasing products. And one approach appears to be winning out.

Jon Keegan20h

73%

Rani Molla22h

Here’s another sign Anthropic’s enterprise tools are killing it: the AI firm now captures 73% of all spending among companies buying AI tools for the first time, Axios reports, citing data from Ramp, a fintech company that provides corporate cards and expense management software. That’s up from 50% in January, when it was tied with OpenAI.

As we’ve noted, Big Tech is pivoting from experimentation to revenue — and enterprise is where that shift is playing out.