Winamp will use Deezer technology to launch a premium streaming service in 2027, combining subscription music with local files, radio, podcasts and cloud libraries. It sounds useful, but the important details remain missing.

Before Spotify, Apple Music and recommendation algorithms decided what everyone should hear next, millions of listeners stored their music on a computer. Those collections were assembled from ripped CDs, purchased downloads and MP3 files acquired through methods best left between the listener and their internet service provider.

Winamp was the software that made those libraries manageable.

The Windows music player let users organize playlists, adjust an equalizer, install plug-ins, run visualizations and completely change the interface through downloadable skins. It was fast, flexible and considerably more personal than most of the streaming apps that eventually replaced it.

Now Winamp wants another opportunity to shape digital music playback.

The company announced on July 29, 2026, that it has formed a strategic partnership with Deezer to develop a new Winamp Player. Scheduled for the first half of 2027, the software will combine a Winamp-branded premium streaming service with locally stored music, internet radio, podcasts, personal cloud collections and other audio sources.

That is considerably more interesting than launching yet another streaming service. The world already has enough of those, along with enough annual price increases to make renting the same albums indefinitely feel like a retirement strategy.

What Was Winamp?

Winamp first appeared in 1997 and was developed by Justin Frankel through his company Nullsoft. It arrived as the MP3 format was beginning to change how music could be stored, shared and played on personal computers.

Winamp did not invent digital music, but it helped make large computer-based libraries practical for ordinary users. AOL acquired Nullsoft in 1999 after the software had attracted millions of listeners.

Its popularity was about more than playback. Winamp gave users control over how their music was organized and how the software looked. Thousands of skins and plug-ins turned it into something that felt personal rather than another corporate storefront.

The player later drifted through ownership changes and uncertain development. AOL announced plans to discontinue it in 2013, although the shutdown did not happen as expected. Belgian internet radio company Radionomy acquired Winamp and Shoutcast in 2014.

Winamp never completely disappeared. A legacy Windows version remains available, while newer mobile applications have been released for iOS and Android.

Calling the 2027 product a full comeback is therefore not entirely accurate. Winamp has been trying to rebuild itself for several years. The significant change is the addition of a complete subscription streaming service powered by Deezer.

What Are Winamp and Deezer Building?

Deezer will supply its catalog and white-label streaming technology, but the service will be presented as a native Winamp subscription rather than Deezer wearing someone else’s jacket.

Winamp says the new player will bring together streaming music, local files, internet radio, podcasts and personal cloud libraries within one customizable interface. The company is also promising social features and new tools for music discovery and organization.

The existing desktop player will eventually gain access to the premium subscription.

Winamp and Deezer also claim that more than 40 million people still actively use the desktop software worldwide. That is an extraordinary figure for an application many listeners assumed had disappeared alongside dial-up modems, although neither company has provided independent verification.

Deezer is a logical infrastructure partner. It already operates a global streaming service and supplies music technology to telecommunications companies, hardware manufacturers and other third-party businesses.

For Deezer, Winamp creates another way to distribute its catalog without having to persuade Spotify and Apple Music subscribers to switch directly to the Deezer brand.

Why Deezer?

Deezer enters the Winamp partnership from a stronger financial position than its smaller market share might suggest.

The French streaming company reported revenue of €268.2 million for the first half of 2026, adjusted EBITDA of €8.5 million and net income of €6.7 million, reversing a €7.6 million loss from the same period one year earlier. Direct subscribers increased 8.7% to 5.8 million, although Deezer’s total subscriber base declined 3.5% to 8.9 million as subscriptions supplied through outside partnerships continued to fall.

Deezer is not remotely close to challenging Spotify or Apple Music on scale, but it has a licensed global catalog, lossless streaming technology and an expanding business supplying music infrastructure to other companies. Winamp does not need Deezer to become the world’s largest streaming platform. It needs Deezer to make the plumbing work.

What Are the Alternatives to Winamp?

Winamp will not be the first platform to combine locally stored music with subscription streaming.

Roon is the most obvious alternative for audiophiles. It merges local files with services including TIDAL and Qobuz, adds extensive metadata and discovery tools, supports bit-perfect playback and works with more than 1,000 compatible audio devices. It is also priced at $14.99 monthly, $149.88 annually or $829.99 for a lifetime subscription, before anyone buys a dedicated Roon server.

Audirvāna Studio also combines local files stored on computers, USB drives and NAS devices with TIDAL, Qobuz, Presto Music, HIGHRESAUDIO and internet radio. Its emphasis is more firmly placed on high-resolution playback and computer-based audio than mainstream discovery or social features.

Plexamp is another strong option for people who primarily want to organize and stream their own music collection. It offers gapless playback, loudness leveling, smart playlists, CarPlay, Android Auto, AirPlay and Chromecast, but it does not merge that library with a full subscription music catalog in the same way Winamp is proposing.

Winamp’s opportunity is therefore not proving that local and streamed music can coexist. Roon and Audirvāna already do that rather well. Its challenge is making the experience simpler, cheaper and more accessible to people who do not own a dedicated server or use “bit-perfect signal path” in casual conversation.

Why Does This Matter?

Streaming services have made access to music remarkably convenient, but they have also separated many listeners from the music they already own.

Someone with 2,000 ripped CDs, purchased FLAC downloads, independent releases, live recordings and carefully organized playlists often needs one application for local files and another for streaming. Internet radio, podcasts and cloud storage may require additional software.

Winamp wants to put those sources inside one player.

That could be genuinely useful because streaming and locally owned music serve different purposes. Streaming is excellent for discovery and immediate access. Local files offer permanence, specific masterings and control over recordings that may disappear when licensing agreements change.

Winamp is not the first company to attempt this. Roon combines local libraries with several streaming platforms, while music servers from companies such as Innuos merge stored files with subscription catalogs. Apple Music also allows listeners to manage local music alongside its streaming library.

The difference is that those solutions are either tied to specific ecosystems, require another subscription or work best with dedicated audio hardware.

A properly designed Winamp application could offer a broader software-based alternative for listeners who want one library without purchasing a dedicated music server.

The word “could” is carrying a rather large loudspeaker up the stairs.

Who Would Use It?

The most obvious audience consists of listeners who built substantial digital music libraries during the CD-ripping and download eras and never deleted them simply because Spotify arrived.

That includes people with MP3, AAC, ALAC, WAV and FLAC collections stored on computers, external drives or cloud accounts. It also includes listeners who buy music directly from artists, collect alternate masterings or own recordings unavailable through mainstream services.

Existing Winamp users are another natural audience, assuming that 40 million figure is remotely close to reality.

The player could also appeal to people who dislike the increasingly uniform design of major streaming applications. Customization was central to Winamp’s original appeal, and the company says users will again be able to modify layouts and create skins.

Audiophiles should remain cautious.

Deezer currently offers lossless streaming, but neither company has confirmed that the Winamp subscription will include lossless audio. There is also no information about bit-perfect playback, exclusive audio modes, automatic sample-rate switching, USB DAC support or compatibility with network streamers.

Until those details emerge, this is a promising library-management concept rather than an audiophile playback platform.

What We Don’t Know

Winamp has not announced the price, launch countries or which operating systems will be supported at launch. There are no confirmed details about streaming quality, offline downloads, family plans, playlist transfers or whether existing Deezer subscribers will have any upgrade path.

The company has not identified which cloud-storage platforms will work or whether the player will index NAS drives and shared network folders.

Hardware support is another large blank space. Winamp has not confirmed AirPlay, Google Cast, UPnP, DLNA, smart-speaker integration, automotive support or direct playback through connected audio components.

It has also not demonstrated the finished software.

At this stage, Winamp and Deezer have announced a partnership, a concept and a 2027 release window. That is not the same thing as delivering a stable player capable of merging a complicated local library with millions of streamed tracks.

The Bottom Line

Winamp and Deezer are pursuing an idea that deserves attention.

Major streaming services have spent years improving discovery and recommendation systems while often treating locally owned music as something listeners abandoned with CD binders and beige computers.

Millions of people still own large digital collections and should not require a separate archaeological expedition every time they want to play them.

Deezer gives Winamp a catalog, licensing relationships and streaming infrastructure. Winamp brings a recognizable brand and decades of experience with local music playback and customization.

That combination could produce a genuinely useful alternative to conventional streaming apps. It could also become another Deezer-powered service with a different logo, several customizable buttons and a marketing department leaning very heavily on the past.

The difference will come down to pricing, audio quality, hardware support and whether Winamp can combine local and streamed music without turning library management into unpaid technical support.

The llama may still whip the llama’s ass.

It also needs to support your DAC.

For more information: winamp.com

Related Reading:

We can all see how the newest digital products require flexibility and the ability to scale quickly together with the business. At the same time, they must cope with growing workloads. That’s why teams choose is*hosting for their Linux VPS hosting to get the job done. Flexible server solutions make it possible to configure resources, select a suitable location, and deploy the environment automatically.

The list of users who will find this helpful is extensive: DevOps engineers, web application developers, SaaS platforms, and many others. Using open-source tools helps them create a reliable, transparent, and scalable environment for development and operations.

DevOps is today one of the key approaches to creating and managing software products. It connects development, testing, and operations, helping teams work faster. Open-source tools make workflows easier and flexible. Open source is basically built to bend. You can customize your tools, stitch them together, and automate the boring stuff. Honestly, it’s a perfect fit for pretty much anything — web, mobile, cloud, data, or AI. 

It is fair to say that several specific DevOps components help automate processes, improve system stability, and shorten the time between product development and launch. These include:

• application build automation;
• containerization of production services;
• server configuration management;
• infrastructure health monitoring;
• secure storage of credentials.

Less manual grunt work means fewer mistakes. Auto-deploys let you ship updates in a flash, while smart monitoring catches performance bottlenecks before they become fires. 

Scalable Infrastructure for DevOps Teams

In practice, any DevOps model requires a reliable and powerful infrastructure. Every project, even the smallest one, must have the ability to quickly scale resources as more and more users join or as technical requirements change. The server environment must support flexible power settings, rapid creation of new environments, and stable service operation.

International IaaS providers offer VPS/VDS, dedicated servers, VPN services, and additional infrastructure solutions. Users can configure key server parameters, including:

• processing power;
• amount of RAM;
• drive type;
• data center location.

A broad infrastructure geography makes it possible to select suitable locations for different projects and reduce access latency. Such opportunities are in demand in various fields. Online stores use scalable servers to handle large numbers of requests.

When choosing DevOps infrastructure, it is important to consider not only current challenges but also growth prospects. A flexible server platform should allow configuration changes without complex migrations. Among the most popular features are:

• setting up CPU and RAM;
• use of fast NVMe drives;
• support for different operating systems;
• connecting additional IP addresses;
• automatic deployment of environments;
• integration via API tools.

This model makes it easier for technicians and helps launch new projects faster. DevOps teams can create test environments, run experiments, and migrate applications between environments with minimal time investment.

Fundamentals of System Security and Resilience

As soon as digital products begin to evolve, securing the system as a whole becomes just as important as maintaining performance. Good companies must secure the users data, provide uninterrupted service, and minimize the risk of breakdowns.

It is exactly these advanced open-source tools that help not only manage the entire system but also monitor it clearly. Teams will be able to track changes, audit settings, and implement additional security measures. When building a secure DevOps environment, special attention is paid to the following aspects:

• isolation of virtual servers;
• data backup;
• protecting network connections;
• use of IPv4 and IPv6;
• user access control.

KVM technology ensures effective isolation of virtual machines and stable operation of VPS servers. SSDs and NVMe drives enable faster processing, and the latest generation of Intel Xeon and AMD EPYC processors allows you to run resource-intensive applications.

How to Choose Infrastructure for Long-Term Development

Choose a server that actually grows with you. You want to scale smoothly without hitting hidden fees. Look for clear, pay-as-you-go pricing — it’s the smartest way for startups to get top-tier tech without breaking the bank. 

Flexible servers work for almost any project. Building an edtech app? Marketing tools? Analytics? AI? To summarize, good infrastructure is the secret sauce to keep your product growing smoothly.

When you need a comprehensive strategy, keep in mind that it can only be based on a successful combination of elements. This includes automation, open-source tools, and high-quality server infrastructure. Together, they create a secure and scalable DevOps environment. These technologies help developers build products faster, monitor all critical processes, and manage resources as efficiently as possible.

Photo credit: Crissa Graves
Crissa Graves has spent years making the old Game Boy Camera feel less like a novelty and more like something you might actually carry. Her latest prototype, the GBD-M2, takes that work further. It starts with a fully working Game Boy Pocket and turns the whole thing into a dedicated handheld camera with a removable camera module and interchangeable CS-mount lenses.

Video of Camera GBD-M2 in action

Once I’ve reached a comfortable place with the project, I plan on sharing the files to build yourself

[image or embed]

— Game Boy Camera (@gameboycamera.com) July 28, 2026 at 4:20 PM

You still get the same 128 by 112 pixel grayscale photographs that the first Game Boy Camera produced in 1998. The small little sensor from that era remains at the heart of every photo, but Graves has simply given it a body that makes you feel like you’re holding a proper camera while still being able to play all of the original Game Boy titles.

There is a dedicated shutter button that sits exactly where your finger would normally go, and the screen is a nice illuminated IPS LCD that you can see outside, which is a great benefit. An 1800mAh battery will keep you snapping away for hours, and you can even charge it via USB-C. The connection connector remains, allowing you to transfer your images to a Game Boy Printer. Surprisingly, Graves claims that you can even play some casual games like Tetris or Pokémon one-handed, despite the fact that the speaker has been removed from this version.

Graves had already completed various iterations of this Camera M project, demonstrating that the whole concept was feasible, although those older versions required hacking into a Game Boy Pocket and had some rather improvised power boards that you’d have to cobble together. The GBD-M2 is a completely new design that takes the Game Boy’s original CPU and RAM and transplants them onto a fresh new circuit board without interfering with a functional handheld. The layout is more camera-like, with bespoke buttons and a shell that has been reduced to a more manageable size after several rounds of design improvement. You also get two battery options: a LiPo battery that can be charged by USB-C, or a set of simple AA cells that you can replace out when they run out.

The camera module detaches, allowing you to replace the custom CS-mount version with alternative lenses, insert the original Game Boy Camera, or simply load a typical old game cartridge. Graves has already created a separate camera module the size of a regular game cartridge, using an iPhone lens and a custom board, which still works in any Game Boy. The GBD-M2 employs the same technique, but in a body designed to be portable and usable.

There are no mass production plans for this device yet, but Graves has mentioned releasing both DIY kits and a limited run of the finished model. You can access all of the files and updates on her GitHub page for the Camera M2 project. For the time being, the GBD-M2 is only a functional prototype, the latest stage in a long journey to find new methods to revitalize an old relic from 1998. You may monitor her progress at gameboycamera.com and on her Bluesky account.
[Source]

Apple has reached 1.5 billion paid subscriptions across its platform, adding another milestone to an Apple Services business that generated record June-quarter revenue of more than $30 billion.

CEO Tim Cook disclosed the milestone during Apple’s fiscal third-quarter earnings report. Apple didn’t provide a breakdown showing how many subscriptions belong to its own services or third-party apps sold through the App Store.

The figure represents paid subscriptions rather than 1.5 billion individual customers. A single customer can hold multiple subscriptions, and multiple devices in use, across services such as iCloud, Apple Music, Apple TV, and third-party apps.

Apple’s Services division generated $30.74 billion during the June quarter, up 12% from a year earlier and setting a new June-quarter record. Revenue increased across advertising, the App Store, AppleCare, music, video, iCloud, and payment services.

The result still came in below the $31.22 billion analysts expected. Apple cited foreign exchange headwinds, but the shortfall against Wall Street’s estimate doesn’t change the division’s year-over-year growth or its new June-quarter record.

Apple Services revenue

Services generated $30.9 billion in the previous quarter, continuing the division’s steady growth beyond hardware sales. The latest result rose from $27.4 billion in the same quarter a year earlier.

Apple also has more than 2.5 billion active devices in use, giving the company a large installed base for subscriptions and other digital services. Paid and transacting accounts reached records during the previous quarter, although Apple didn’t disclose a comparable subscription count at the time.

The 1.5 billion total is more than 50% higher than the nearly one billion paid subscriptions Apple reported in 2023. The increase shows continued expansion across Apple’s subscription ecosystem even as quarterly Services revenue missed analysts’ forecast.

Services have become a major source of repeat business beyond Apple’s hardware sales. The latest milestone reinforces that growth, while the Wall Street miss shows investors are still judging the division by the revenue those subscriptions produce.

“It’s not red,” IBM’s design center manager patiently explained. “It’s magenta.”

The safety function officer wasn’t buying it. Between his fingers, he held up a small, rubbery disc that looked like a pencil eraser. “You know it’s red and I know it’s red,” he said.

Amazon linked multiple high-profile open-source software supply chain attacks targeting the Node Package Manager (npm) ecosystem to North Korean hackers.

The cloud computing giant linked the compromises of the typo-crypto, debug, chalk, and axios libraries to the Sapphire Sleet threat actor, also known as BlueNoroff and Stardust Chollima.

Initial activity started with trojanizing the typo-crypto package in March 2025, which Amazon believes served as a testing ground. It then escalated in September of the same year with the compromise of the widely used debug and chalk packages, affecting an estimated 10% of cloud environments within two hours.

In March 2026, the hacker targeted axios, one of npm’s most popular packages with over 100 million weekly downloads.

It should be noted that the axios incident has already been publicly attributed to DPRK-linked actors, but Amazon connected it to the earlier package compromises.

Amazon says the attacker gained access by socially engineering package maintainers, and then published malicious updates that were automatically distributed to unsuspecting users.

The attribution to Sapphire Sleet has medium confidence and is based on the shared tactics, techniques, and procedures (TTPs) observed in the campaign, the command-and-control (C2) infrastructure, and various operational similarities.

Also, the researchers believe the attacker had a financial motivation, targeting popular packages to gain indirect access to a large pool of potential downstream victims at once.

Amazon also highlights several trends that have emerged from the recent supply-chain attacks:

• Attackers are splitting malicious functionality across multiple seemingly benign packages, making detection more difficult.
• Threat actors are spending months building trust by maintaining legitimate projects or becoming contributors before introducing malicious code.
• Malicious behavior is increasingly decoupled from package contents, relying on external scripts, configuration files, or servers that can be weaponized later.
• Malware is using stronger encryption and multi-stage payloads, with runtime or remotely fetched keys that hinder static analysis.
• Payloads are becoming environment-aware, delaying execution unless they detect real developer or production environments to evade analysis sandboxes.
• Attackers are increasingly exploiting “slopsquatting” by registering package names hallucinated by AI coding assistants, hoping developers or autonomous coding agents will install them.

Many of these tactics are enhanced and simplified by AI, Amazon explains, as they help attackers generate code, documentation, and maintainer identities.

Amazon highlighted a multi-faceted response to these dangers, including reporting its findings and intelligence to the community, collaborating with OpenSSF and other industry partners, and investing $12.5M in the Akrites initiative, which helps protect critical open-source software from AI-enabled attacks.

Security teams log 54% of successful attacks and alert on just 14%. The rest move through your environment unseen.

The Picus whitepaper shows how breach and attack simulation tests your SIEM and EDR rules so threats stop slipping by detection.

Get the whitepaper

from the funny-how-that-works dept

Just recently Karl warned that we were going to see some absolute nonsense as the US sought to somehow “ban” Chinese AI models from being used in the US. That seems to already be happening. It kicked off with talk that the US might “fight Chinese AI” using nearly identical arguments to what was used to ban (or force the sale of) TikTok before it. Some combination of “national security threat” combined with “oh no China” propaganda.

But most of the AI industry is now speaking out, in an open letter put together by Nvidia, against the potential path that the Trump administration considered taking: an attempt to ban or limit so-called “open weight” models. The companies seem to recognize that focusing on holding back these competitive models would actually do much more damage to the wider AI ecosystem.

There were notable exceptions from the campaign in defense of open weight models: Anthropic, OpenAI, and Google (the three leading frontier model labs) were not initially signed onto the letter. Though their absence quickly became the story — leading OpenAI and Google to reconsider and sign onto the letter days after it came out.

That left one major player off the letter: Anthropic (a company that has so far refused to release any open weight models). And now the company is trying to explain itself, but seems to only be digging itself a deeper hole.

One of the problems here is that the leading Chinese AI models tend to be open weight models, which can be downloaded and run locally, as compared to the leading frontier models from US companies which require you to access them via their own hosted models. Yes, most of the leading Chinese models also offer (sometimes significantly cheaper) cloud/API access to their models, but you can also run them yourself (for the smaller models directly on your own computers, or for the larger models via your own cloud setup).

There’s no inherent reason why the best open weight models are coming out of China, other than that they seem to have recognized that it may be the best way to get more people to use them and to compete against the American frontier models, which are much more proprietary and locked up. The strategy is a recognition that offering a compelling, more open alternative is how to get people to adopt your system over the American frontier models. If a generation of developers builds on top of Kimi or Qwen or one of the other Chinese open weight models, they become the de facto infrastructure for the next generation of digital tools.

In the same manner that Linux quietly became the substrate of the open internet, and it’s likely that an open weight model may become the equivalent for the next generation. Organizations may rely on frontier models for really deep work, but so much can be done with open weight models that a winner here becomes the commodity infrastructure provider for a new generation of software. That’s why any proposed restrictions on open weight models would get everything precisely backwards. It would guarantee that the wider open ecosystem gets built on non-American tools. Yet, the discussion around such bans seems to treat these as just another software product, rather than a fight over how the infrastructure of the internet will work going forward.

Of course, that’s not the only argument the Trump admin is using to try to stop these models. Last week they focused on claims that Kimi’s K3 model (the latest model to shake up the US market, despite being not quite as good as the frontier models) must have been “distilled” from Anthropic’s Fable 5.

“If we see, especially that overseas models are stealing from our great companies, we have the ability to sanction them because of this theft,” Bessent told Fox Business’ “Mornings with Maria” on Tuesday.

Bessent said the technical term for this theft is called distillation, which is an AI training method where a smaller, less capable model is built using outputs from an existing, stronger model. Anthropic sent a letter to the U.S. Senate Committee on Banking, Housing, and Urban Affairs last month alleging that the Chinese tech company Alibaba had carried out the “the largest known distillation attack” against it to date.

This is rich for a variety of reasons, not the least of which is that all of the frontier AI models were built by feeding their training models whatever information they could get their hands on, including (in Anthropic’s case) building a pirate library of downloaded books for which it had to pay out a pretty massive settlement to authors.

Distillation is not quite the same thing, but is functionally similar. It’s taking the work of an existing model to fine tune the model you’re working on. The claims about Kimi K3 seem somewhat exaggerated, as the initial claims were that it was distilled based on Anthropic’s Fable 5 release, but multiple people I’ve spoken to don’t see how that’s possible, given how Fable 5 has only been out for a little while (and then was turned off for a while due to the US government freaking out over nothing).

No matter what, distilled models are likely to be less powerful, and at least a decent period behind the frontier models, given that they’ll need access to the frontier models and time to train based on them. There’s also some dispute over how the open weight models may be using distillation, and which part of the training process works best.

But either way, the freakout over distillation seems… ridiculous. Bessent calling it “theft” is nonsense. Just as training a model on copyrighted works is a form of reading (which shouldn’t implicate copyright in the first place), so too is distilling, which is (in effect) training your model by having it compare its initial answers to similar answers from a frontier model and then adjusting based on the different results. It’s a form of learning based on observed results by others, not “stealing.” Pretending that it’s stealing or somehow should face sanctions or other consequences will put US AI development in a bad, bad spot.

Which brings us back to that letter. Here’s the case it actually makes:

Open weights also strengthen competition and competition is what keeps the gains of AI broadly shared rather than concentrated in a few hands. By allowing many organizations to build, adapt, and deploy advanced models, open weights create rivalry not only among model developers but across cloud chips, applications, and services. That competition spurs innovation, drives down costs, and distributes the benefits of AI broadly across our economy.

Open weights also give customers greater control. As organizations invest in AI, they want to know that they will not become locked into a single provider or lose the knowledge and capabilities they build over time. Open weight models help provide that assurance by allowing organizations to control their own data, evaluate and adapt models to their own needs, and deploy them wherever their business requirements demand. And as organizations create value with AI, open weights allow them to own that value through self-improving models, specialized capabilities, and accumulated knowledge that drive American sovereignty and prosperity.

The letter is exactly correct. I’ve talked about the importance of open-weight and local models for taking back control over the open web and making sure that we don’t run a repeat of what the earlier internet had of a few giant companies taking over the web.

Of course, Anthropic (which hasn’t released any open weight models) was conspicuously absent from the signatory block of that letter. Earlier this week, Anthropic’s Dario Amodei came out and tried to explain/justify the company’s stance, which boils down to: “we don’t think anyone should ban open weight models… but we do think the US should ban all the conditions that make quality open weight models possible.”

Amodei argues that simply banning Chinese open weight models wouldn’t solve the alleged “threats” that people are concerned about, though he admits directly that it would act as protectionist industrial policy that could benefit American AI companies (like Anthropic):

But banning the use of these models by US businesses does nothing to address this risk, because bad actors are unlikely to be legitimate US businesses. It would protect US AI companies from competition, but that has never been my goal.

It feels a bit like he’s protesting too much regarding the protectionism here, while trying to have it both ways. He claims he really has the best interests of safety at hand, and is against protectionist ideas, but it’s hard to square that with the rest of the article.

While he says the US shouldn’t ban open weight models (and it shouldn’t), he then puts a bunch of conditions on it, which would make it that much more difficult for the current crop of open weight models to compete. Namely, he leans in on the Sinophobia that has become popular these days in warning about “CCP” influence over models (which… should be less of a concern with open weight models, since those who use versions not hosted by the Chinese companies can adjust the models to deal with those concerns).

But then he says that we should punish Chinese AI companies for engaging in distillation:

We should crack down on industrial-scale distillation operations. Distillation is a much more compute-efficient process than training models from scratch. It allows China to build much better models than its number of chips would ordinarily enable, and thus partially evade chip bans. Distillation does not allow the CCP to obtain equivalent or superior AI capabilities to the US, but it can bring the Chinese frontier to within a few months of the US frontier. It is true that many of the companies carrying out these operations release open-weights models—but the open weights are far less relevant than the fact that the operations are backed by an authoritarian state seeking to overtake the US at the frontier. We should have policy interventions to deter this behavior. A blanket ban on open-weights models is neither the correct remedy nor something we have called for.

To be fair to Amodei, not everything on his list is competitor-hobbling. He also wants chip export controls tightened (a policy that predates this fight and has its own problems, but at least isn’t aimed at a business model), and he wants mandatory pre-release safety testing for all sufficiently capable models — open or closed, foreign or domestic, Claude included. That last one is the tell, though, and not in the way he intends: if you genuinely believe capability-based testing is the right lever, and you’ve just said bans “would protect US AI companies from competition, but that has never been my goal,” then what is the argument about distillation doing on the list at all? Testing catches dangerous capabilities regardless of how the model got them. The distillation crackdown adds nothing on safety. It only serves to kneecap cheaper competition.

And even the “safety testing” plank isn’t as neutral as it sounds. While safety testing is obviously important, when legally mandated, it can quickly turn into an expensive compliance-function of box-checking that only the largest companies can do, taking us back to the world of just a few providers, and limiting smaller competitive models from really being viable. While there are legitimate reasons for it, it can also create its own moat.

The proposed crackdown on distillation is just asking the state to step in and block lower-cost competitors from competing. Yes, these models can be competitive, but they should be driving the leading frontier models to continue to improve and to provide more value. What Amodei is asking for here is basically the US government to help prevent lower cost, lower quality competitors from pushing the floor of the AI market upwards.

Now, to be clear, as with any technology, you can claim that a more open, more widely available, more powerful version can be misused. But that has always been the case and we, in the US, have tended to default to allowing the technology to proceed, and figuring out ways to minimize the dangers/increase the good uses, rather than resorting to assuming the tech will be abused and working backwards to block all possible abuses. Historically, seeking to pre-vet technologies tends not to work well, and (often) opens up the market to foreign competitors to simply build better products.

The open letter makes a sharper version of this point, and you can see why Anthropic wouldn’t want to put its name to this point in particular:

Relying solely on closed models is not inherently safe: they can be breached, misused, or fail in ways that outsiders cannot detect. And concentrating advanced AI capabilities behind a small number of closed models compounds that risk. It results in a small number of single points of failure, weakens competition, and leaves critical technology in the hands of a few providers. Open weight models, on the other hand, allow a broad community of researchers and developers to examine their behavior, identify vulnerabilities, develop safeguards, and improve them over time. Just as open-source software demonstrated that transparency can be more secure than obscurity, AI safety may depend on giving more people the ability to test and strengthen the models on which society relies. It allows for rigorous benchmarking and evaluation, red teaming, and protections tied to real and demonstrated harms rather than assuming that closed systems are safer by default.

Amodei also claims he supports the general argument of the open letter, but he disagrees with the idea that open weight models lead to better security:

This brings me to the open letter. I agree with much of it: open weights expand access to the AI economy, they strengthen competition at least for some use cases, and they give customers greater control. Concerns about distillation should be addressed through targeted legal and commercial frameworks—the same measure I described above. But I don’t agree with the letter’s assertions that open-weights models necessarily make it easier to develop safeguards or that broad access to capabilities necessarily helps defenders more than attackers. It seems at least as likely to me that the opposite will be true.

This strikes me as a repeat of the age-old fight that always shows up in discussions of open source technologies: the claim that by making them open, security vulnerabilities are easier to find. Of course, what we’ve seen historically in other spaces is that this actually means that security vulnerabilities are more quickly patched, rather than in the “security by obscurity” space, where they can remain open (and possibly exploited) for much longer.

Amodei is asserting that the AI space is somehow different, though without much evidence for that other than what feels like a bit of fear-mongering about “weaponizing pandemic-level viruses.”

Of course, part of the problem here is that it often feels like Anthropic treats “crying wolf” as a marketing strategy, whereby much of the company is focused on talking up “our tools are soooooooo dangerous that you need us in there to protect you from them.” Even if there’s some truth to it, it’s awfully convenient that the same argument also happens to justify banning, punishing, or limiting the cheaper, more open, more user-controllable alternatives.

In the end, the federal government still might try to punish the Chinese open models in some form or another just because they view current American industrial policy in very nationalistic terms. But that won’t be good for the wider ecosystem, or for the general incentives to innovate. And, worst of all, it makes it that much harder to build a world where we’re not entirely dependent on a few giant companies controlling the “brains” of the tools the rest of us rely on.

Filed Under: ai, ai safety, competition, dario amodei, distillation, frontier models, industrial policy, open weights

Companies: alibaba, anthropic, google, kimi, nvidia, openai

UL’s Prof Martin Hayes on his systems theory research, why ‘pi-shaped’ graduates are the future of engineering, and the importance of patience.

Prof Martin Hayes is a professor of digital technologies at University of Limerick (UL) who describes his research as sitting “within the space of systems theory for machine learning and AI”.

Hayes’ research looks at how to manage system resources intelligently when they’re subject to uncertainty or mixed messaging introduced by communication channels, sensors or human operators. A large part of his work focuses on the “robust” performance of AI in safety-critical environments.

“How do we correctly choose settings when an AI system’s outputs feed into decision-making by people, without getting that handoff wrong?” asks Hayes.

“Health is an obvious area where such solutions have to be correct 100pc of the time.

“The alternative has real consequences in terms of negative outcomes – so I’m asking how we can guarantee a health system that will exploit the benefits of AI while working optimally for every citizen every time.”

Hayes tells SiliconRepublic.com that as AI and machine learning technologies move from research labs into safety-critical, high-stakes settings, these systems “can’t simply be accurate on the average”, and that understanding how to build in basic levels of robustness allows technology to be deployed responsibly in regulated domains such as digital health.

“Professionals working in the health system, be they medics, engineers or administrators, need to understand not just how to use AI tools, but how to trust, interrogate, explain and govern them appropriately,” he says. “Without that grounding, adoption either stalls through excessive caution or accelerates without the safeguards that safety-critical settings demand.

“Translational research that provides a foundation for those who are already working in or who wish to participate in the new European Health Data Space is a key focus of my work.”

‘Pi-shaped graduates’

With his work spanning from systems theory to engineering and the real-world domains where these tools get deployed, Hayes says that the interdisciplinary nature of the research is actually a very rewarding part of the job.

In particular, he enjoys the collaborative side of it where he works directly with industry partners, clinicians and SMEs to “understand where the genuine skills gaps are and then translating that into education and research that actually closes them”.

However, according to Hayes the most satisfying part of the job is seeing graduates go on to responsibly deploying this understanding and thinking in the workplace.

“I believe strongly that the future of engineering education revolves around the growth of such ‘pi-shaped’ graduates who have the basic skills in AI-enabled data engineering but who also have the necessary allied health skills to be able to apply those solutions in a safe, human-centred fashion,” he says.

Hayes has worn and continues to wear many hats at UL, where he has worked since 1997.

He is the academic lead for the UL@Work Human Capital Initiative project, which aims to develop digital, industry 4.0 talent through flexible, innovative, technology-enabled, experiential learning. Hayes says his involvement in the UL@Work project has been “hugely insightful”.

“One key takeaway is that universities need to work together and collaborative programmes like Digital Europe are essential in enabling institutions to pool their resources so that they can offer students the bespoke learning that fits their individual needs.”

Hayes is also collaborating with various European universities as principal investigator for a number of Digital Europe projects – such as that of Sustainable Healthcare with Digital Health Data Competence (SUSA).

SUSA is a €12.4m Digital Europe-funded project led by the University of Oulu in Finland that aims to close the digital skills gap in European healthcare and support the EU’s Digital Decade and European Health Data Space ambitions.

The project aims to deliver revamped bachelor’s, master’s and standalone lifelong-learning modules, built around 20 shared SUSA learning objectives that have been benchmarked against frameworks such as the WHO Digital Health Competence Framework.

Hayes is UL’s principal investigator for SUSA, leading the ‘Workpackage’, which frames and co-designs SUSA activities in order to maximise impact.

“We lead in two specific tasks: investigating how to best deliver education on the optimal use of advanced digital technologies in health – particularly XR/AR and digital twin technology – and designing the SUSA employer framework that connects students with industry most efficiently,” he explains.

“UL’s contribution to the SUSA digital ecosystem draws on existing UL@Work advisory board models and Skillnet partnerships to keep the curriculum grounded in current workplace needs.”

Fundamentals and patience

As someone working in such a future-focused research area, we asked Hayes about what advice he might have for someone considering a career similar to his own.

First thing on the list? “Build a strong foundation in the fundamentals,” he says.

By fundamentals, he means systems theory, mathematics and statistics.

“These are what let you adapt, configure and ultimately deploy solutions as the technology moves on,” he explains.

Next, he advises newcomers to seek out interdisciplinary collaboration early and not to be “afraid to work at the boundary between engineering and the domains where it gets applied, whether that’s healthcare, manufacturing or elsewhere”.

“Get involved in industry-facing projects where you can; the most consistent feedback we get from our students is that they’ve always enjoyed it most when they’ve been exposed to real-world constraints either through UL’s co-op education programme or the in-house projects they complete during their studies,” Hayes adds.

“Ultimately this sharpens the R&D questions you ask and makes you a more valuable resource,” he says.

“Finally, be patient. Your career, much like a trustworthy AI system, is built over the long term and will inevitably require you to actively manage many uncertain situations.

“Embracing that challenge will give you the confidence to achieve your goals. Repetition builds competence!”

Don’t miss out on the knowledge you need to succeed. Sign up for the Daily Brief, Silicon Republic’s digest of need-to-know sci-tech news.

Xiaomi stepped into the large family SUV game on July 30 with the SkyNomad N90 Max, an extended-range electric vehicle built around the idea that a car can serve as living space first and transportation second. Presales opened the same day in China at 299,900 yuan, roughly 44,000 dollars, with official deliveries set for September. The company also showed a smaller five-seat N70 Max sibling starting lower, but the seven-seat N90 Max is the clear centerpiece.

The N90 Max measures 5285 mm in length, 1998 mm in width, and 1825 mm in height, with a wheelbase of 3080 mm. With a completely flat floor and good long seat rails, the groundwork is created for Xiaomi’s Kunlun architecture, a platform designed expressly for this type of customizable cabin rather than driving for driving itself.

Sale

Anker SOLIX C1000 Gen 2 Portable Power Station, 2,000W (Peak 3,000W) Solar Generator, Full Charge in…

• 49 Min UltraFast Recharging: With upgraded HyperFlash tech, fully recharge at 1,600W—for outage prepping, camping trips, or tailgating events…
• 2,000W Output via 10 Ports: Delivers 2,000W (3,000W peak) and 1,024Wh capacity. Power up to 10 devices—ideal for emergency backup, remote work…
• Compact and Portable: Easily carry, store, and move from room to room, your RV, or even on beach and park outings. C1000 Gen 2 is 14% smaller and…

The N90 Max delivers a powerful punch, thanks to a dual-motor all-wheel-drive system that produces a decent 310 kW (416 horsepower). The rear motor produces 210 kW, and the front one adds another 100 kW to the mix. It takes a decent 5.9 seconds to accelerate from 0 to 100 kilometers per hour. You are limited to 190 km/h, but 0-100 in under 6 seconds gives you an excellent idea of the N90 Max’s capability. A 1.5-liter turbocharged engine from Harbin Dongan serves as a generator, producing 112 kW and charging CALB’s 76 kWh ternary NMC battery. With the battery empty, the N90 Max has a pure-electric range of 464 km on the CLTC cycle. If you put some gas in the tank, you’ll have a total distance of 1705 kilometers (1059 miles). When the battery becomes low, the WLTC cycle consumes approximately 6.26 liters of petrol every 100 kilometers. A 15-minute DC charge adds 285 kilometers to your electric range, while charging from 20 to 80% takes less than 18 minutes.

The N90 Max features a 2+2+3 layout. When parked, the front seats will power swivel fully 180 degrees, allowing you and your passengers to face the second row. If you choose the center table, it will slide out of the tunnel, creating a living space up to 1370 mm wide in four-person mode. Second-row seats include zero-gravity designs with footrests and a 16-point massaging system, providing plenty of extra comfort. You also get a 9-liter compressor fridge in the center island, which has cup and phone slots. The N90 Max has some serious sound setup going on, with up to 25 speakers outputting 4,890 watts over a 7.1.4 arrangement. The rear seats feature a 21.4-inch 3K screen, while the driver has a 20-inch head-up display, an 8.8-inch instrument cluster, and a 16.1-inch primary screen. The majority of seat and table adjustments are controlled by voice commands.

Xiaomi describes the parked cabin as a studio for one, a café for two, a lounge for three, or a playroom for the entire family, so it is quite adaptable. The N90 Max Camping Edition exhibits the same flexibility, with a pop-up roof, a rooftop bed platform, tent attachment points, side cabinets, and an optional removable table. This version is suited for five persons and has the same powertrain as the normal N90 Max. The car includes a roof-mounted LiDAR unit as well as a rear-facing solid-state device with an accuracy of a few centimeters. This is combined with some extremely strong computers, courtesy of NVIDIA’s Thor-level technology, as well as Xiaomi’s whole HAD assisted driving package, which includes their XLA cognitive model. The vehicle’s structure is meant to last, with a frame made of high-strength steel and aluminum.

The three rows are connected together by a continuous hot formed door ring that runs the entire width of the vehicle. Underneath, there’s a really innovative suspension system that combines double wishbone front geometry with a multilink rear configuration, and the vehicle can handle a decent soak with a water wading depth of 750mm. It also has enough electricity to keep your electronics charged, with 220 volt charging outlets and 6.6kW vehicle-to-load capability, making it excellent for camping trips or emergency circumstances. The range extender itself operates silently enough that Xiaomi claims the cabin remains as quiet as a pure electric vehicle even when it is powered up. The pre-sales price for the flagship Sky Nomad N90 Max is 299,900 yuan ($44,322), while the somewhat smaller N70 Max is 259,900 yuan ($38,410).