The cost of genome sequencing is falling far ahead of Moore’s Law, and a San Diego company now has it down to $100. CC-licensed photo by Dunk đ on Flickr.
You can sign up to receive each dayâs Start Up post by email. Youâll need to click a confirmation link, so no spam.
A selection of 9 links for you. Based. (Friday was on Saturday.) I’m @charlesarthur on Twitter. On Threads: charles_arthur. On Mastodon: https://newsie.social/@charlesarthur. On Bluesky: @charlesarthur.bsky.social. Observations and links welcome.
Amazon service was taken down by AI coding bot âą Financial Times
Rafe Rosner-Uddin:
»
Amazonâs cloud unit has suffered at least two outages due to errors involving its own AI tools, leading some employees to raise doubts about the US tech giantâs push to roll out these coding assistants.
Amazon Web Services experienced a 13-hour interruption to one system used by its customers in mid-December after engineers allowed its Kiro AI coding tool to make certain changes, according to four people familiar with the matter.
The people said the agentic tool, which can take autonomous actions on behalf of users, determined that the best course of action was to âdelete and recreate the environmentâ.
Amazon posted an internal postmortem about the âoutageâ of the AWS system, which lets customers explore the costs of its services. Multiple Amazon employees told the FT that this was the second occasion in recent months in which one of the groupâs AI tools had been at the centre of a service disruption.
âWeâve already seen at least two production outages [in the past few months],â said one senior AWS employee. âThe engineers let the AI [agent] resolve an issue without intervention. The outages were small but entirely foreseeable.â
AWS, which accounts for 60% of Amazonâs operating profits, is seeking to build and deploy AI tools including âagentsâ capable of taking actions independently based on human instructions.
âŠAmazon said it was a âcoincidence that AI tools were involvedâ and that âthe same issue could occur with any developer tool or manual actionâ.
âIn both instances, this was user error, not AI error,â Amazon said, adding that it had not seen evidence that mistakes were more common with AI tools. The company said the incident in December was an âextremely limited eventâ affecting only a single service in parts of mainland China. Amazon added that the second incident did not have an impact on a âcustomer facing AWS serviceâ.
«
The TV series Silicon Valley was, as ever, remarkable prescient. This scene (with an AI called “Son of Anton”), was written and filmed years ago. The scene also, coincidentally, forecasts – in its last seconds – what happened when Joanna Stern at the WSJ let a chatbot run their vending machine.
unique link to this extract
Suspect in Tumbler Ridge school shooting described violent scenarios to ChatGPT âą The Verge
Terrence O’Brien:
»
The suspect in the mass shooting at Tumbler Ridge, British Columbia, Jesse Van Rootselaar, was raising alarms among employees at OpenAI months before the shooting took place. This past June, Jesse had conversations with ChatGPT involving descriptions of gun violence that triggered the chatbotâs automated review system. Several employees raised concerns that her posts could be a precursor to real-world violence and encouraged company leaders to contact the authorities, but they ultimately declined.
OpenAI spokesperson Kayla Wood told The Verge that, while the company considered referring the account to law enforcement, it was ultimately decided that it did not constitute an âimminent and credible riskâ of harm to others. Wood said that a review of the logs did not indicate there was active or imminent planning of violence. The company banned Rootselaarâs account, but it does not appear to have taken any further precautionary action.
Wood said, âOur thoughts are with everyone affected by the Tumbler Ridge tragedy. We proactively reached out to the Royal Canadian Mounted Police with information on the individual and their use of ChatGPT, and weâll continue to support their investigation.â
«
So 1) there’s a mechanism for OpenAI to be alerted when people are having violent conversations 2) OpenAI feels it’s in a position to decide what is and isn’t notable. “Several” employees raised concerns? Let’s see how this gets fed back up the management chain there.
unique link to this extract
Enthusiasts used their home computers to search for ET. Now scientists are homing in on 100 signals they found âą Phys.org
Robert Sanders, University of California:
»
For 21 years, between 1999 and 2020, millions of people worldwide loaned UC Berkeley scientists their computers to search for signs of advanced civilizations in our galaxy.
The projectâcalled SETI@home, after the Search for Extraterrestrial Intelligence (SETI)âgenerated a loyal following eager to participate in one of the most popular crowd-sourced projects in the early days of the internet. They downloaded the SETI@home software to their home computers and allowed it to analyze data recorded at the now-defunct Arecibo Observatory in Puerto Rico to find unusual radio signals from space.
All told, these computations produced 12 billion detectionsâ”momentary blips of energy at a particular frequency coming from a particular point in the sky,” according to computer scientist and project co-founder David Anderson.
After 10 years of work, the SETI@home team has now finished analyzing those detections, winnowing them down to about a million “candidate” signals and then to 100 that are worth a second look. They have been pointing China’s Five-hundred-meter Aperture Spherical Telescope, a radio telescope referred to as FAST, at these targets since July, hoping to see the signals again.
Though the FAST data are not yet analyzed, Anderson admits he doesn’t expect to find a signal from ET. But the results of the SETI@home projectâpresented in two papers (first, second) published last year in The Astronomical Journalâprovide lessons for future searches and point to potential flaws in ongoing searches.
“If we don’t find ET, what we can say is that we established a new sensitivity level. If there were a signal above a certain power, we would have found it,” he said. “Some of our conclusions are that the project didn’t completely work the way we thought it was going to. And we have a long list of things that we would have done differently and that future sky survey projects should do differently.”
«
Pluribus: delayed for now.
unique link to this extract
PromptSpy ushers in the era of Android threats using GenAI âą We Live Security
Lukas Stefanko:
»
ESET researchers uncovered the first known case of Android malware abusing generative AI for context-aware user interface manipulation. While machine learning has been used to similar ends already â just recently, researchers at Dr.WEB found Android.Phantom, which uses TensorFlow machine learning models to analyze advertisement screenshots and automatically click on detected elements for large scale ad fraud â this is the first time we have seen generative AI deployed in this manner.
Because the attackers rely on prompting an AI model (in this instance, Googleâs Gemini) to guide malicious UI manipulation, we have named this family PromptSpy. This is the second AI powered malware we have discovered â following PromptLock in August 2025, the first known case of AI-driven ransomware.
While generative AI is deployed only in a relatively minor part of PromptSpy’s code â that responsible for achieving persistence â it still has a significant impact on the malware’s adaptability. Specifically, Gemini is used to analyze the current screen and provide PromptSpy with step-by-step instructions on how to ensure the malicious app remains pinned in the recent apps list, thus preventing it from being easily swiped away or killed by the system.
The AI model and prompt are predefined in the code and cannot be changed. Since Android malware often relies on UI navigation, leveraging generative AI enables the threat actors to adapt to more or less any device, layout, or OS version, which can greatly expand the pool of potential victims.
«
*dogcoffeefire.gif*
unique link to this extract
US farmers are rejecting multimillion-dollar datacentre bids for their land: âIâm not for saleâ âą The Guardian
Niamh Rowe:
»
When two men knocked on Ida Huddlestonâs door last May, they carried a contract worth more than $33m in exchange for the Kentucky farm that had fed her family for centuries.
According to Huddleston, the menâs client, an unnamed âFortune 100 companyâ, sought her 650 acres (260 hectares) in Mason county for an unspecified industrial development. Finding out any more would require signing a non-disclosure agreement.
More than a dozen of her neighbors received the same knock. Searching public records for answers, they discovered that a new customer had applied for a 2.2 gigawatt project from the local power plant, nearly double its annual generation capacity. The unknown company was building a datacentre.
âYou donât have enough to buy me out. Iâm not for sale. Leave me alone, Iâm satisfied,â Huddleston, 82, later told the men.
As tech companies race to build the massive datacentres needed to power artificial intelligence across the US and the world, bids like the one for Huddlestonâs land are appearing on rural doorsteps nationwide. Globally, 40,000 acres of powered land â real estate prepped for datacentre development â are projected to be needed for new projects over the next five years, double the amount currently in use.
Yet despite sums that often dwarf the landâs recent value, farmers are increasingly shutting the door. At least five of Huddlestonâs neighbors gave similar categorical rejections, including one who was told he could name any price.
In Pennsylvania, a farmer rejected $15m in January for land heâd worked for 50 years. A Wisconsin farmer turned down $80m the same month. Other landowners have declined offers exceeding $120,000 per acre â prices unimaginable just a few years ago.
«
In the TV adaptation, this would become a standoff – mediated by heavies – between the two opposing sides. If money won’t work, how about a little, shall we say, persuasion? (In the TV series Justified, also set in Kentucky, the men trying to buy up the land want to turn it into marijuana farms, legal or illegal. They get persuasive about it.)
A side note: since The Guardian recognises my IP well enough to badger me about visits, could it add some Javascript to spell “centre” (and other words) the British way rather than the American one, to save me editing it? (I might have to add something to the script that compiles the links.)
unique link to this extract
OpenClaw-fuelled ordering frenzy creates Apple Mac shortage âą Tom’s Hardware
Jowi Morales:
»
Some Apple customers have recently been taken by surprise by the order lead times on several Mac models with upgraded Unified Memory quotas, which could be largely driven by the immense popularity of OpenClaw, the locally-run open-source AI agent thatâs taking the internet by storm and sending users scrambling for Macs to run the AI.
While you can still get base model units of the MacBook Air, iMac, M4 Mac mini, and other basic models on the same day, upgrading memory can now increase delivery wait times by up to three weeks.
However, going for the highest possible memory capacity on high-end models greatly increases your waiting time, with the M3 Ultra Mac Studio with 512GB of Unified Memory taking up to five to six weeks to be delivered.
âŠWhile data centres are hungry for AI GPUs and some startups are using multi-gaming GPU setups to train AI models, theyâre not ideal for personal agentic AI run locally. This is especially true if you use a huge 70-billion parameter model in FP16 for your agent, which would require around 140GB of memory just for weights, according to AI investor Ben Pouladian. That means that it wouldnât fit inside a single RTX 5090 with 32GB of VRAM, and even if you manage to connect five graphics cards for a total of 160GB of memory, youâre still bound by the PCIe bottleneck.
Appleâs Unified Memory architecture (UMA) fixes that problem. Even though HBM is still way faster than the LPDDR used in Macs and MacBooks, the fact that all the processing units â CPU, GPU, and NPU â share the same memory means that they donât have to deal with PCIe bottlenecks or require technologies similar to NVLink, which is typically only found on data centre-class graphics cards.
«
Itâs called the âFitbit for Fartsââand itâs no joke âą WSJ
Christopher Mims:
»
As I type this, a battery-powered computer tucked in my, er, swimsuit area is monitoring for my next flatus. Yes, flatus means âfart.â Donât judge meâitâs for science.
This sensing device, which would have been impossible to make until very recently, sits at the intersection of some of the most important technology trends of the past decade, including miniaturization, continuous monitoring and edgeâas opposed to cloudâcomputing.
And the data itâs designed to collect could help the 40% of U.S. adults whose lives are regularly interrupted by digestive troubles. The frequency and volume of flatulence is a major reason people quit healthy, fiber-rich diets, which can be key to heading off gastrointestinal disorders and colon cancer.
One day, millions of Americans might wear sensors like this one, which I found surprisingly unobtrusive. Playfully referred to by its creators as a âFitbit for farts,â itâs the core of the University of Marylandâs Human Flatus Atlas study. Developed using some of the same components found in smart rings and wireless earbuds, devices like this could eventually share a stage with blood-glucose monitors and heart-monitoring Apple Watches.
âŠGoing from a bulky commercial hydrogen sensor to a smart device about the size of a stack of three nickels was a technical challenge like that of any consumer-health wearable. The team went through 12 iterations before landing on the current design.
âWe had to come up with a completely custom solution,â says Hall. Using a tiny battery required adopting the newer kind of chips found in so-called internet-of-things devices. Users charge the sensor every 24 hours with a custom magnetic charger sourced from Chinaâs smart-ring supply chain. He says such electronics have become far cheaper and more plentiful only recently. âThis would have been impossible three years ago,â he adds.
When the device is charging, it transmits data to the Human Flatus Atlas app on your phone. (Hall says participantsâ data remains confidential.) The sensor sits outside your underwear and can be cleaned with alcohol wipes. Though itâs not waterproof, you can wear it during most workouts. But no cycling!
«
There’s a paper from their earlier studies. The highest number? 175. The lowest? Four. (Thanks Joe S for the link.)
unique link to this extract
The human demotion âą AI Policy Perspectives
Tom Rachman:
»
Exercise might seem pointless if biotech offered a way to instantly make your body healthy and beautiful. Skipping the sweaty workout might not trouble you. But what if future humans would bungle child-rearing when compared with AI nannies, meaning that nurturing your offspring would worsen your kidâs life?
Primitive versions of this dilemma are nearing, like when human drivers endanger lives when compared with self-driving cars. âHuman in the loopâ could flip from a safety promise to a threat. Meritocracy would mean that no humans need apply.
The bookworm economist Tyler Cowen cites people as the great obstacle to explosive AI growth. During a public event, he pointed at the audience, smiling toward the human âbottlenecksâ before him. âHere they are: bottleneck, bottleneck. Hi, good to see you! And some of you are terrified. You are going to be even bigger bottlenecks,â he said. âBut my goodness, once it starts changing what the world looks like, there will be much more opposition. Not necessarily on what Iâd call doomster grounds. But people [saying], like: âHey, I see this has benefits, but I grew up, trained my kids to live in some other kind of world. I donât want this!â And thatâs going to be a massive fight.â
The most agonizing aspect of our demotion could be social, once someone prefers a machine to you. Youâre seeing precursors every time family members opt to gaze at a screen rather than gaze at you. We blame smartphones, and social media, and the adolescent brain.
But wait till your spouse jilts you for a personified agent. That rejection may feel unbearable: you canât compete anymore. And once your loved ones prefer AI companions, you might seek them for yourself, spreading the social downgrade of our kind.
Already, the dread is becoming political, with odd alliances forming among right-wing politicos, liberal artsy types and religious traditionalists, united in horror at an imagined future of disempowered humanity, stripped of dignity, obsolete. You can imagine tomorrowâs political opportunist, eyeing a dejected crowd of humans before him, and thundering: âHow dare they?!â
Will he mean the machines?
«
There is a lot of philosophising going on about quite where AI is going to land us. So it helps to read as many perspectives as possible.
unique link to this extract
Scrappy San Diego startup goes toe-to-toe with gene-sequencing giant Illumina âą San Diego Union-Tribune
Noelle Harff:
»
Element Biosciences is going toe-to-toe with gene-sequencing giant Illumina, unveiling a device that can read DNA for half the price of the industry leaderâs technology.
On Thursday, Element Biosciences announced that its high-throughput benchtop sequencing device called VITARI can deliver a whole genome for $100.
A few years ago, Illumina came out with the NovaSeq, which turned heads for its $200 high-throughput whole genome sequencer. Elementâs VITARI is not only half the price of the NovaSeq, but itâs also a fraction of the size while maintaining lab-grade readings, the company said.
âWe hope to push the field as a whole and raise the bar,â said Matthew Kellinger, co-founder and vice president of biochemistry at Element. âWe were seeing sequencing stagnate in the time before Element came around. There was no competition. There was one model. Thatâs how it was.â
Element was founded in 2017 after three high-ranking Illumina employees walked out: Molly He is the CEO; Michael Previte is the chief technology officer of advanced research; and Kellinger is the vice president of biochemistry.
Today, Element and Illumina are in the midst of a legal battle. Element alleges antitrust and unfair competition, while Illumina alleges patent infringement. Both sides are pursuing claims in federal courts.
âŠElementâs new instrument is marketed toward biotech firms, large pharmaceutical companies and academic centers that require substantial sequencing capacity.
While Illumina and Element are marketing whole genome sequencing at $200 and $100, respectively, most labs are sequencing genomes for roughly $400 per genome, said Jepsen.
«
The first human genome sequence, in 2000, cost hundreds of millions of dollars. The fall in cost since has been faster than Moore’s Law would predict, perhaps driven by the huge volume and demand.
unique link to this extract
| âą Why do social networks drive us a little mad? âą Why does angry content seem to dominate what we see? âą How much of a role do algorithms play in affecting what we see and do online? âą What can we do about it? âą Did Facebook have any inkling of what was coming in Myanmar in 2016? Read Social Warming, my latest book, and find answers – and more. |
Errata, corrigenda and ai no corrida: A copy-pasting error meant that Microsoft’s new system for writing on glass was said to take 10 to 15 seconds per bit, rather than 10^-15 (a millionth billionth) of a second per bit. Different!
Thanks to the many people who got in touch about the WordPress hassles which made Friday’s edition turn into Saturday’s. The glitch was at WordPress’s end, as I thought, so it was a question of waiting. If the problem repeats and persists, I’ll probably figure out a new location at Substack, but with any luck that won’t happen before Overspill 3,000, due in about five years.
The Overspill: when there's more that I want to say 