Neuralink Brain Malfunction Is Good News
Hello from high above Colorado,
I write this on airplane WiFi. A marvel of modern engineering. In fact, earlier I was texting from an airplane to my friend Andrew Heaton who was using Starling satellite internet to get his texts in the middle of a forest. It's the future friends.
I've got a contrary take on the Neuralink story today plus my usual cord cutting take on the Disney-Warner streaming bundle.
Hold on, they’re asking me if I want coffee.
Enjoy!
Tom
Big Story
"Neuralink says its first in-human brain implant encounters problem"
Neuralink gets a lot more press than other Brain-Computer Interface companies who are farther along. I'm always quick to note that when they tout a success and I think it's only fair to note that when they have a setback. These are setbacks other companies have been through as well.
Neuralink reported it had implanted its tech in the first human patient as a trial. The company posted Wednesday that some of the threads that connected the device to neurons retracted reducing the ability to measure speed and accuracy. Neuralink compensated by modifying the algorithm that interpreted the signals and improving the techniques for translating those signals to cursor movements.
Also really important to note, the implant is not posing a health risk. It's just a connectivity issues. And in some ways, this may be a good way to improve the algorithm to make ti more sensitive. The patient is reportedly still using it at least 8 hours a day.
More Stories
"Disney and Warner bundle streaming services to rival Netflix"
Disney and Warner Brothers Discovery announced they plan to offer their streaming services as a bundle. You'll be able to get Disney+, Hulu and Max for one price starting sometime this summer. The companies didn't announce a price for the combo but presumably it will come at some discount from getting them separately. ESPN+ was not mentioned as part of the bundle, but Bob Iger previously said main ESPN content would be coming to Disney+ by the end of the year.
This is going to happen more often. If you've been listening to me on Cordkillers, DTNS or reading this newsletter, you know I've been saying we're in the consolidation phase of streaming. Companies did their best to acquire viewers, now they are trying to turn those audiences into businesses. As they do that some will have to team up with others. That may take the form of actual company mergers and acquisitions. Paramount is pretty much up for sale with Skydance and Sony with Apollo Ventures interested.
But it's also going to take the form of bundles like Max and Disney products. Disney is also teaming up with Fox and Warner on a combined sports service. And apparently Paramount also has a plan to bundle Paramount+Showtime with Peacock as well.
I know the fun joke is to say, "it's becoming cable again." It's not. You don't have one company providing you all the viewing. Bundles are just ways to share audiences and will have the effect of bringing down the cost a little for those who would want all the bundled services anyway. There isn't just one bundle but multiples. It's also not cable because the bundle isn't your only option. You can still get the individual services and control your costs.
I expect we'll end up with about 2-5 streaming services available in various combinations. I'm going to guess that will take 3-5 years.
"OpenAI offers a peek behind the curtain of its AI's secret instructions | TechCrunch"
"ChatGPT maker OpenAI exploring how to 'responsibly' make AI erotica"
OpenAI published its "model spec" which lists the rules that govern its models like ChatGPT. These are used to create filters and tuning that reduce false or dangerous responses. Developer intent is the prime directive. So if the developer using a model wants it to only answer questions about language learning, it won't stray to other topics.
The instructions also describe how OpenAI draws the line between health information and NSFW content. And the company is always re-evaluating its guidelines. The document also says, "We're exploring whether we can responsibly provide the ability to generate NSFW content in age-appropriate contexts." Joanne Jang, an OpenAI model lead told NPR that the company wants to start a conversation about where that line should be drawn.
It looks to me like this is a responsible way to go about things. Draw the line the best you can. Err on the side of safety. But keep thinking about it and talk to folks in the community about what they think. It would all be easier if it was an actual open-source project but given that it's not, this is pretty good policy in my opinion.
"TikTok is adding an ‘AI-generated’ label to watermarked third-party content - The Verge"
TikTok will start adding an "AI-generated" label to images that are watermarked and made by third parties. This is in addition to labeling images generated by TikTok's own tools. It will start by supporting Adobe's Content Credentials watermark. It will also detect content made by its own tools that were downloaded and re-uploaded. This won't catch everything, but it will make it clear that some images are generated and for folks acting responsibly, identify what is and is not. If you want this to be a perfect defense against AI manipulation you don't have it. But it will normalize labeling generated content.
"DeepMind adds a diffusion engine to latest protein-folding software | Ars Technica"
DeepMInd AlphaFold version 3 has added some functions to the model that can find protein folding. Protein folding is important to understanding functions and can lead to drug discoveries. Previous versions of AlphaFold could find the structure but the new version includes various ways the protein interacts with other proteins, DNA, chemicals, membranes which is a big part of telling what it's good for.
We talked about this on DTNS Wednesday, but I didn't realize at the time that DeepMind was using a diffusion engine to do the interaction side of the results. It trains the diffusion model on two levels of noise so it can learn to tell functional structure. It took 20,000 instances to get test structures right 97% of the time and 60,000 to get protein-protein interfaces right.
As we said on DTNS, it's. Big time saver. AlphaFold saved months to years of trying out folding structures and this can save weeks to months of tracking down what interactions they have.
A lot of folks don't want their data used to train Large Language Models. So, when Stack Overflow announced a partnership with OpenAI to license its data for training, a lot of folks on Stack Overflow began deleting their questions and answers to prevent it from being used in training. You can question whether that's an overreaction, but it's certainly within someone's rights to delete their posts for whatever reason.
Of course, the platform also has the right to kick you off if you do that, and in some cases that's what Stack Overflow moderators have been doing. It's not the best move from a customer relations standpoint though. You're probably not going to win over hearts and minds by kicking people off the platform.
For Context
"A crushing backlash to Apple’s new iPad ad | Ars Technica"
We talked about this a lot on Good Day Internet Wednesday. What I didn't consider at the time was all the memes about crushing things with a hydraulic press. Thanks to the multiple people who pointed that out. This is Apple's producers living in a bubble. To them, it was a nod to a funny meme. To anyone who doesn't know the meme, which is a lot of people, it looked odd. That's the internet for you.
Geekbench results are leaking out and the top line is that it's faster than skeptics expected.
"WiiM announces two new wireless audiophile streamers — one with a touchscreen - The Verge"
Just a reminder that there are high-end devices for those who really want the best sound from streaming music.
"Alibaba rolls out latest version of its large language model"
Here's the important bit, "The latest [Alibaba Cloud] model fares better than OpenAI’s GPT-4 model in language and creation capabilities, but fell short of its competitor in other categories like knowledge, reasoning and math, according to OpenCompass." I'm pretty sure this is why the US is placing so many restrictions on Chinese tech. The US does not want to give China the edge in AI.
"Roblox cuts annual bookings forecast on muted player spending | Reuters"
The service that is a candidate to one day evolve into something you might call the metaverse is seeing reduced spending. It used to be just a cash river, so it's probably going to be fine, but interesting to note.
Interesting Reads
"Huawei’s Kirin 9010 Is a Reality Check for China’s Semiconductor Ambitions | iFixit News"
"LinkedIn’s new daily games are surprisingly fun to play - The Verge"
"Inside Microsoft’s Xbox turmoil - The Verge"
"Apple’s Next CEO: List of AAPL Insiders Who Could Succeed Tim Cook - Bloomberg"