Security

Another day, another Windows exploit code

Nightmare Eclipse, the prolific zero-day vulnerability hunter with an axe to grind against Microsoft, released yet another exploit late Wednesday that the researcher claims will spawn a command prompt that provides total access to the BitLocker volume.

This bug, called GreatXML, was “an accidental discovery,” according to the researcher, who said it only took four hours to find. They claim this exploit (published on GitHub and Git-based code-hosting platforms) can bypass BitLocker on any system that has ever run a Microsoft Defender Offline scan at any point in the past.

GreatXML comes just a day after Nightmare released exploit code for RoguePlanet, which allows local privilege escalation and leads to SYSTEM-level control over an affected machine. This brings the researcher’s zero-day count to eight. The earlier six – RedSun, UnDefend, BlueHammer, YellowKey, GreenPlasma, and MiniPlasma – all have patches as of this week’s Patch Tuesday event. 

Redmond on Wednesday told The Register that it is aware of RoguePlanet, and “actively investigating the validity and potential applicability of these claims.” The Windows giant didn’t immediately respond to our inquiries about GreatXML, including when it planned to issue a patch.

Microsoft has said none of the vulnerabilities were reported via its official channels prior to being made public. The company also banned Nightmare’s earlier GitHub account, and seemingly threatened legal action before dialing back its rhetoric after steep backlash from the security community.

Nightmare Eclipse, who some researchers suggest is an ex-Microsoft employee, harbors a very personal grudge against the Windows giant and its communications with bug hunters. They have promised to keep the zero-days coming, but waffle on the timing. 

Last month, the researcher pledged a big July 14 drop: “I will make sure your bones are shattered that day,” and then added, “nothing will be released this June (or maybe I will release smtg, depending on circumstances).”

On Tuesday, they changed course. “I will be unable to mass disclose zerodays in July 14th, RoguePlanet took way more time than expected and truly drained me. I might take a break but I can’t say for sure what I will be doing for next month, maybe it’s nothing, maybe it’s smtg.”

A day later, Nightmare released the “accidental” GreatXML BitLocker bypass. 

According to the researcher, the BitLocker bypass first requires copying “unattend.xml” and the “Recovery” directory to the root of the recovery partition. The next step is rebooting into WinRE by Shift-clicking Restart. “If everything was done correctly, a shell with unrestricted access to the bitlocker volume will spawn,” Nightmare wrote.

Also, if the scan hasn’t even been initiated on the Windows system, first you’d need to either log in and initiate it, or “figure out a way to boot into WinRE in offline scan state.”

Security sleuth Will Dormann followed Nightmare’s steps to reproduce GreatXML, and said the writeup seems “flawed.” In his testing, Dormann said the command prompt appeared the next time a Defender Offline scan ran.

“And in order to trigger a Microsoft Defender Offline scan, you both need to be logged in to Windows, and also have admin credentials,” he wrote on social media. “And if you’ve already got that level of access, you can just turn off bitlocker.”

“The writeup for GreatXML suggests that the prerequisite is that Windows Defender Offline has been executed at some point in the past,” Dormann added. “And that after planting two files in WinRE, all you need to do is [Shift]-reboot into WinRE, and Windows will automatically go into Microsoft Defender Offline scan mode. But this is not the case in any of the 3 lineages of Win11 that I have handy.” ®

from the the-end-of-ten-blue-links dept

Google didn’t invent full-text search of the Internet – that honor belongs to early pioneers such as WebCrawler, Lycos and AltaVista. But for the last 25 years or so, Google has been synonymous with online searching, providing the quickest and most effective way to find things online (although its results may be getting worse.) More recently, it has been adding to its search engine more features based on generative AI, first with its AI Overviews in 2024, and then a year later with its AI Mode in Search. Now it has announced the latest stage in that evolution with what it calls “A new era for AI Search”:

It’s more intuitive than ever, dynamically expanding to give you space to describe exactly what you need. Designed to anticipate your intent, it also helps you formulate your question with AI-powered suggestions that go beyond autocomplete. And you can search across modalities, using text, images, files, videos or Chrome tabs as inputs.

This new incarnation effectively turns search into a chatbot:

You can easily ask a follow-up question right from an AI Overview, and flow into a conversational back and forth with AI Mode. Your context stays with you, and as you explore more deeply, the links and supporting articles get even more relevant. This seamless experience is live today across desktop and mobile, worldwide.

As the the screenshot of the new interface above shows, the traditional search result links that are currently placed under the AI Overview have now been confined to a small panel on the right-hand side of the screen, which shows a cut-down version of today’s list. Users are encouraged to ask follow-up questions from the AI search chatbot, rather than exploring the links themselves.

What this is likely to mean in practice is that even fewer people will follow links to sites, something that was already happening last year; instead, they will engage with Google’s chatbot to gather information indirectly. This is terrible news for access to knowledge because it frames the Google AI search engine as the fount of all knowledge – one that will do all the hard work of finding information and combining it into an easily digested answer that can be interrogated further. It can do that because it has already ingested billions of Web pages and other information sources as part of the Large Language Model (LLM) training process. But search engine users will no longer know what some of those sources are unless they painstakingly click on the links in the new panel.

Most people will not bother, because the AI-generated results will be good enough – or at least will appear to be good enough. Unless visitors to the site take the trouble to follow the links to the sources they won’t really know how reliable those results are. For example, it is possible that the sources are wrong, or misleading; moreover, Google’s LLM may itself introduce new errors and distortions. There is also the question of how Google will insert ads into this AI-generated information, and to what extent advertisers will be able to buy preferential treatment in results.

This new mediated approach is clearly terrible news for Wikipedia – an issue already discussed on Walled Culture earlier this year – and for creators. Google will use the information found in their works, but will not actively encourage people to visit the originals. For many people, summaries will be good enough, and they will never discover the greater riches of the sites and creations that Google’s LLM is based on. Worse still, the original creators such as Wikipedia may not even be mentioned in answers that involve aggregating information from a large number of sources.

Similarly, the new Google search is the publishing industry’s worst nightmare. Not only is Google drawing on material they have published, but it is pushing links to those sources into the background. It seems inevitable that the Web traffic to publishers will fall yet further, making already struggling business models based on advertising even more precarious. That will have knock-on consequences for the funding of many sites – particularly newspapers and magazines – and for the commissioning of work from journalists and other creative professionals. Users won’t even need to visit Google Search much in order to keep up-to-date with topics of interest thanks to Google Search’s new agentic capabilities that will do the work for them in advance:

With information agents, you can stay updated on whatever matters most to you. Your agent will intelligently look across everything on the web, like blogs, news sites and social posts, plus our freshest data, such as real-time info on finance, shopping and sports, to monitor for changes related to your specific question.

In this case, not only will people not visit sites, but the latter will be constantly bombarded by various AI bots seeking information on behalf of users – increasing site running costs, and making sites less usable by humans. Another key announcement from Google will lead to a further flood of agentic activities that will pose new challenges to businesses:

We’re also expanding agentic booking capabilities in Search to a wide range of new tasks, including local experiences and services. Just share your specific criteria — like finding a private karaoke room for six on a Friday night that serves food late — and Search brings together the latest pricing and availability with direct links to finish booking through the provider of your choice. And for select categories like home repair, beauty or pet care, you can ask Google to call businesses on your behalf.

What emerges from Google’s latest announcements is less of a search engine, and more of an immersive virtual environment that is designed to keep people engaging with Google’s services, asking them for information, advice and even delegating actions to them. There is no doubt that many users will find these new features attractive, not least because they can use “conversational voice features” in Gmail, Docs and elsewhere. These are the digital assistants that have been promised for many years, able to understand spoken commands, provide information verbally, and carry out complex operations on behalf of users without the need for any complex training. For many people, that will be a boon, and they will doubtless migrate from the traditional search page, which will still be the default – at least for now – to the latest AI-infused version.

But these impressive technical features come at a high price, even leaving aside issues such as the environmental impact of the huge server farms they require. With the latest incarnation of its search engine, Google is making the World Wide Web as we have known it for over 30 years invisible, and therefore increasingly irrelevant to most people, who will be happy to let Google become their universal user interface to everything. And yet Google still depends on the Internet to supply all the information it is analyzing and repackaging. It risks killing the very thing that sustains it.

There’s another, more subtle issue. The new Google search features make finding information and carrying out actions very easy in many ways. Leaving aside the problem that this will require people to trust what is in effect a huge black box, where the internal workings cannot be examined, with all the loss of control this implies, there is another danger. People who use Google’s powerful new AI search services to offload many of their day-to-day actions may gradually lose the ability to understand the world and to act within it without that constant help. Such a dependence may be great for Google and its advertisers, but it surely cannot be a good thing for the future of society.

Follow me @glynmoody on Mastodon and on Bluesky. Originally published to WalledCulture.

Filed Under: ai, links, open web, search

Companies: google

Here is a constraint that almost no one building physical AI says out loud, even though every one of them is quietly fighting it.

A robot’s intelligence wants three things at once. It wants to be smart, meaning it can reason at the level of a frontier model about an unfamiliar scene. It wants to be fast, meaning it responds inside the tight, deterministic timing a physical control loop demands. And it wants to be free, meaning it keeps working when the network drops, the warehouse Wi-Fi dies, or the machine goes somewhere no signal reaches.

You cannot have all three on one piece of compute. Pick any two.

To be precise, bounded autonomy already works. Industrial arms, drones, and constrained autonomy stacks can be fast and offline because their tasks are narrow. The trilemma bites at the frontier: you cannot put frontier-scale general reasoning, deterministic real-time response, and full offline autonomy into the same power-limited substrate, not for the same control loop.

The 💜 of EU tech

The latest rumblings from the EU tech scene, a story from our wise ol’ founder Boris, and some questionable AI art. It’s free, every week, in your inbox. Sign up now!

A frontier-scale model is smart, and if you stream its sensors to a datacenter it can even be fast, but now it is tethered to a network and no longer free. Shrink that model until it fits on a 15-watt embedded module and it becomes fast and free, but it is no longer frontier-smart. Run the big model in the cloud and query it only occasionally, and you get smart and free, but never fast. Three corners, two available at a time. I have come to think of this as the embodied trilemma, and it is the real reason the edge/cloud question is the hardest architecture decision in robotics. Most teams treat it as a deployment detail. It is closer to a law.

Why you can’t cheat the triangle

The trilemma is not a fashion or a temporary hardware limitation you can wait out. It falls directly out of physics and power budgets.

Frontier reasoning quality currently lives in models that want tens of gigabytes of memory and datacenter-class accelerators. That hardware does not run on a battery a mobile robot can carry. So “smart” forces a choice: either bring the datacenter to the robot through a network link, which sacrifices freedom, or accept a smaller onboard model, which sacrifices smartness.

Real-time control is even less negotiable. A wide-area network round trip adds 30 to 100 milliseconds of latency, and the variance matters more than the average. A control loop that is usually fast but occasionally stalls is worse than one that is reliably mediocre, because controllers are tuned for deterministic timing. The moment “fast” depends on a network, you have surrendered “free,” because the network is now inside your control loop whether you meant it to be or not.

So the triangle holds. Quantization, distillation, and better accelerators move the corners, but they do not collapse them. Anyone claiming otherwise is usually hiding which corner they gave up.

Putting numbers on the triangle

It helps to make the constraint quantitative, because the moment you write the timing down, the corners stop being abstract.

Start with latency. The end-to-end delay of a perception-to-action decision made in the cloud is a sum of terms:

L_cloud = t_capture + t_encode + t_uplink + t_inference + t_downlink + t_decode

Run the same decision onboard and most of that sum disappears:

L_edge = t_capture + t_{inference,local}

The difference between the two is not the inference time, which can actually be lower in the cloud on better hardware. The difference is the network, t_uplink + t_downlink, and more importantly its variance. A measured cloud-robotics setup over a fast wired link saw round trips of roughly 30 milliseconds [7], while real-world deployments commonly sit in the 100 to 300 millisecond range, and wireless links swing far higher. Edge processing, by contrast, pulls round trips down toward 1 to 5 milliseconds because nothing leaves the machine [8].

Now state the rule that decides where a loop can live. A control loop with timing budget L_budget can run on a given compute path only if

L_path + k·σ_jitter ≤ L_budget

where σ_jitter is the standard deviation of the path’s latency and k is the safety factor you need for determinism. That k·σ_jitter term is the quiet killer. Teleoperation studies are blunt about it: a link that holds a steady 100 milliseconds is workable, but one oscillating between 30 and 200 milliseconds produces jerky, unpredictable motion, because the controller cannot plan around delay it cannot predict [9]. The reflex loop’s budget is 1 to 10 milliseconds. No wide-area path satisfies the inequality. The math, not the architect, forbids it.

The table is the argument in one view. Reflex never clears a network round trip. Perception clears it only on unusually good links. Deliberation has budget to spare, which is why it can live in the cloud asynchronously.

Bandwidth closes the case for perception. A single 1080p camera at 30 frames per second produces raw video at 1920 × 1080 × 3 bytes × 30, which is about 1.5 gigabits per second. A modest four-camera plus depth rig clears 6 gigabits per second of raw sensor data. You can compress it, but compression costs latency and the link still has to carry it reliably, everywhere the robot goes. Edge perception is the robotic version of that move. Compress to a semantic representation on the spot; never ship the raw stream.

Finally, the economics, which is just the trilemma with a dollar sign. Onboard compute is a one-time capital cost. Cloud reasoning is an operating cost that accrues with every query:

C_cloud(t) = r·c_token·t

where r is the query rate and c_token the per-token price, against a flat C_edge = C_capex. The two lines cross at t* = C_capex / (r·c_token). Push thirty frames a second to a cloud model and t* arrives almost immediately, so cloud cost dominates the lifetime of the fleet. Route only a few deliberation-class queries per minute upstream and t* recedes over the horizon.

Same hardware, same models, opposite economics, decided entirely by which loop you placed in which corner. The gap is not subtle: a single camera streamed to a cloud vision model at 30 frames per second is on the order of a million inference calls a day per robot, while routing only deliberation-class queries upstream might be a few hundred. Across a fleet, that is the difference between cloud inference being a rounding error and being the largest line on the operating budget.

The escape nobody designed, because biology did it first

Here is the part I find beautiful, and the heart of what I want to argue: the way out of the embodied trilemma is not to solve it. It is to refuse to answer it at a single point.

Your own body is built this way, and it has been for roughly half a billion years.

When you touch a hot stove, your hand pulls back before your brain knows anything happened. That is the spinal reflex arc, a loop that runs through the spinal cord and never waits for the cortex. It is fast and free (it works even if you are barely conscious), and it is emphatically not smart. It does not reason about the stove. It does not need to.

Your retina does something just as telling. It has over a hundred million photoreceptors, but the optic nerve carrying signal to the brain has only about a million fibers [10]. The eye does roughly a hundredfold compression on the spot, locally, before transmitting anything. It does not ship raw pixels up the cable. It ships a processed, compact representation. Fast and free at the edge, by necessity.

And then there is the cortex, which is where the actual reasoning happens. It is slow, it is powerful, and crucially, the body has arranged things so that when the cortex is slow or offline, the reflexes still fire and you still pull your hand back. Evolution put the survival-critical functions where they never depend on the smart, slow part.

That is the whole trick. Biology never built a single neuron that was smart, fast, and free all at once. It built a hierarchy in which different loops each sit at a different corner of the triangle, and it made sure the corner each loop sacrifices is one that loop can afford to lose. Reflexes give up intelligence, which is fine, because a reflex that stops to think is a reflex that gets you killed. The cortex gives up speed, which is fine, because it has been kept off the survival path entirely.

A robot escapes the embodied trilemma the same way, or it does not escape at all.

Mapping the triangle onto a machine

Translate the nervous system into engineering and a practical architecture emerges. A robot has three loops, and each one belongs at a different corner.

The reflex loop (1 to 10 ms): motor control, stabilization, emergency stops. This is the spinal cord. It must be fast and free and is allowed to be dumb. It lives onboard, always, and never touches a network.

The perception loop (30 to 100 ms): detection, tracking, obstacle avoidance, visual odometry, SLAM. This is the retina. It must keep working when the link drops, and the bandwidth math forbids shipping raw sensor data anyway, since even a single camera produces well over a gigabit per second of raw video before compression. So perception compresses at the edge, exactly as the eye does, and emits a compact semantic representation rather than pixels. Fast and free, intelligence traded away on purpose.

The deliberation loop (1 to 10 seconds): task planning, language understanding, deciding what to do when the plan breaks. This is the cortex. It is allowed to be slow, and slowness is exactly the corner it trades away, reaching a frontier model in the cloud asynchronously rather than in the control path. It stays free in the only sense that matters, never holding the robot hostage to a live link. If connectivity vanishes, the robot gets less clever, not less safe.

The interface between these layers is the optic nerve of the system: a deliberately narrow channel carrying detections, tracks, and state summaries, never raw signal. Get that channel right and you have not just an inference boundary. You have defined your logging schema, your training-data pipeline, and your behavior when the link drops, all at once.

The industry is rediscovering the nervous system

What convinces me this is structural, not stylistic, is that the most advanced robotics programs keep reinventing the same hierarchy without necessarily naming it.

Figure AI’s Helix, the system running its humanoid robots through full eight-hour factory shifts, is explicitly two systems: a roughly 7-billion-parameter vision-language model at 7 to 9 Hz for scene understanding and language, coupled to a compact 80-million-parameter visuomotor policy that turns intent into continuous action at 200 Hz [1]. That is cortex and reflex on one robot, a 25-to-1 ratio in update rate between the loop that thinks and the loop that acts, each running at the timescale its job demands. Surveys of edge-cloud collaboration now describe the same division as standard practice, with small onboard models handling real-time perception and privacy-sensitive preprocessing while heavier reasoning is offloaded upstream [4].

Comparisons on real robot data quantify the trade directly: deploying an 11-billion-parameter vision-language model at the network edge held accuracy close to its cloud baseline while shaving only modest latency, whereas a compact 2-billion-parameter model more than halved latency into sub-second territory, paying for the speed with accuracy [5]. Reviews of foundation-model robotics keep flagging the same wall: LLM planners take seconds per decision, fine for the cortex, hopeless for the spinal cord [6]. NVIDIA’s own Jetson deployment guidance reflects it too, with optimized onboard inference for perception and policy and larger models living upstream [2].

Different teams, different machines, the same triangle, the same corners. When that many independent efforts converge, you are looking at structure, not style.

Lessons from the ultimate airgap

The starkest place to watch the trilemma bite is underwater robotics. An ROV below the surface has effectively no real-time link to the cloud. The ocean is the ultimate airgap, the freedom corner taken to its absolute extreme. In hands-on underwater robotics builds, perception (detection and tracking, optimized with TensorRT) runs entirely on an onboard module, while language-level mission interaction and fleet reasoning reach a frontier model in the cloud only asynchronously, on surfaced or relayed data, and never inside a control loop. The architecture is not a preference there. The water enforces it.

Three principles follow, and they generalize far beyond the sea.

Design for the disconnected case first. If the robot is safe and useful with zero connectivity, the cloud becomes pure upside: better reasoning, fleet learning, human oversight. If the robot needs the cloud to stay safe, you have built a cortex with no spinal cord, a liability on wheels.

Treat the narrow channel as a contract, not a cable. The compressed representation crossing the edge/cloud boundary is the single most important interface in the system. Teams that treat it as an afterthought re-architect twice.

Remember the trilemma is also an economics statement. Onboard compute is paid once, at purchase. Cloud reasoning is paid forever, per token. Routing only deliberation-class queries upstream, a few per minute instead of thirty frames per second, changes fleet unit economics by orders of magnitude. Cloud-inference cost can quietly become the largest operating line on a robotics program that put the wrong loop in the wrong corner.

The corners will move. The triangle won’t.

Onboard modules get more capable every generation, and distillation keeps narrowing the gap between edge models and their cloud teachers. Early-exit inference, where confident predictions resolve locally and only hard cases escalate, is maturing fast [3][5]. The deliberation loop will migrate partly onboard over the next few years, especially for safety-relevant replanning. The corners of the triangle will keep sliding.

But the triangle itself does not go away, because it is anchored in physics and energy, not in any model generation. Smart, fast, and free will never coexist on a single substrate as long as frontier intelligence costs more power than a robot can carry and the speed of light caps how fast a remote answer can return. The teams that internalize this, and that consciously assign each loop the corner it can afford to lose, will ship robots that work when the network does not. The rest will keep learning, in the field and at the worst possible moment, that they accidentally wired their spinal cord through a datacenter.

Evolution settled this argument before there were spines. We are just catching up.

Context windows are becoming a computational bottleneck. The longer an agent runs, the more tokens accumulate from retrieved documents, reasoning traces and conversation history, and the more memory and compute that growing context demands. Most existing solutions either degrade model accuracy, require the full context to load before compression begins, or produce memory savings that don’t translate into real speedups in standard serving infrastructure.

A research team from NYU, Columbia, Princeton, University of Maryland, Harvard and Lawrence Livermore National Laboratory published a paper this week that proposes a novel fix. The researchers introduce the concept of  Latent Context Language Models, or LCLMs, a family of encoder-decoder compression models that compress input context before it reaches the decoder. The models are open-sourced on HuggingFace.

Unlike KV cache compression methods — the dominant approach in the field, which still materialize the full KV cache before evicting entries — LCLMs compress the input token sequence before decoder prefill, so higher compression ratios directly reduce decoder-side compute and memory. The paper reports LCLMs at 16x compression produced output 8.8 times faster than KV cache baselines on the RULER long-context benchmark.

“These ballooning contexts take up memory and compute, and they are becoming a computational bottleneck for LLMs,” Micah Goldblum, co-lead advisor on the project and a researcher at Columbia University, told VentureBeat. “Our goal was to train language models end-to-end that can handle very long contexts efficiently and accurately. If you can make such a language model, everything becomes cheaper and faster.”

What LCLMs can do

LCLMs let models process much longer contexts than would otherwise be practical, at a fraction of the memory and compute cost, without the accuracy degradation that makes most compression methods a poor tradeoff in production.

At 4x compression, the paper reports accuracy of 91.76% on the RULER benchmark, compared to 94.41% with no compression at all. That is less than a 3 point drop for cutting context to a quarter of its original size. At 16x compression, where 93.75% of input tokens are removed, accuracy fell to 75.06%. Every KV cache method tested at the same compression ratio scored lower.

The gains hold on shorter inputs too. On GSM8K math word problems, where the full prompt is compressed rather than just retrieved documents, LCLMs outscored every other method tested regardless of compression ratio.

How it was built

The architecture pairs a 0.6B encoder with a 4B decoder. The encoder compresses blocks of input tokens into shorter sequences of latent embeddings. The decoder processes those in place of the original tokens. Training ran across more than 350 billion tokens.

The training recipe mixes three data types:

• Continual pre-training data with compressed and uncompressed spans interleaved throughout

• Supervised fine-tuning data covering reasoning and long-context tasks

• An auxiliary reconstruction task that pushes the encoder to retain fine-grained detail

The combination addresses a tradeoff that limited earlier compression work, where preserving reconstruction accuracy came at the cost of general task performance.

An architecture search identified the optimal configuration. The paper found that scaling the decoder matters more than scaling the encoder.

Where it fits in an agentic stack

An LCLM is not an abstract research concept. It is designed to work with an existing stack. “You can simply swap out LCLMs for any existing LLM,” Goldblum said. “Whenever you retrieve data such as documents and want to dump it into your model’s context, simply run those documents through the LCLM’s compressor first.”

He noted that in the research paper, the researchers demonstrated how to build agents that selectively decompress useful text. 

“Think about this like a human skimming content before zooming in on relevant details,” Goldblum said.

Goldblum also cautioned that teams integrating the approach into existing agentic pipelines will need to tune their RAG systems accordingly.

“We also haven’t worked on online compression of reasoning traces,” he said. “The naive approach of just occasionally compressing the trace while generating it might work, but that remains to be determined.”

What this means for enterprises

Context windows are growing faster than inference infrastructure can keep up, and enterprises are already spending to fix it. VB Pulse Q1 2026 survey data from 100-plus employee organizations shows hybrid retrieval adoption intent tripling from 10.3% in January to 33.3% in March. Retrieval optimization overtook evaluation as the top investment priority by March, reaching 28.9% of qualified respondents.

Three things stand out for teams evaluating production fit:

1. Inference cost scales with context length. At 1 million tokens, uncompressed inference with standard KV cache methods runs out of memory on a single H200 GPU. The paper reports LCLMs at 16x compression remain within memory bounds at that context length.

2. RAG pipeline integration requires tuning. Teams with existing RAG pipelines will need to validate compression behavior against their retrieval quality metrics before deploying at scale.

3. Reasoning trace compression is unsolved. For agents running long reasoning chains, context growth from the trace is a separate problem from document retrieval. Goldblum acknowledged the gap directly: the naive approach of periodic trace compression might work but has not been tested.

The models are available at huggingface.co/latent-context and the code at github.com/LeonLixyz/LCLM.

“The biggest things our architectures do is give your model access to much larger contexts, but they also unlock multiscale approaches where your model can skim vast amounts of text or code super fast and then only zooms in and fully reads a small portion of the most useful text,” Goldblum said.

Meta on Wednesday previewed upcoming additions to its video-editing app Edits at an invite-only creator event in L.A., showing off features like a new AI assistant and a desktop version of the previously mobile-only app.

The company also announced other new tools will launch in the app today, such as a “Beta” tab for experiments and expanded audience insights.

Edits first arrived last year as a direct competitor to ByteDance’s CapCut. With the addition of the new and upcoming tools, Meta is looking to both retain and attract new users.

The upcoming AI assistant will help creators analyze their insights and brainstorm ideas for their content. The assistant will use their Instagram data, like their views and video-retention insights, to help them see what’s working and why. It will suggest video ideas based on performance and suggest making content with trending audio.

By integrating an AI assistant directly into Edits, Meta is aiming to keep creators engaged on Instagram as it continues to compete with TikTok and YouTube for creators’ attention. Additionally, by offering creators content ideas, Meta is encouraging more frequent posting, which could, in turn, boost user engagement. Direct access to an AI assistant also gets rid of the need for creators to turn to outside tools like ChatGPT when brainstorming content ideas and understanding performance.

Meta launched a similar AI assistant tool for creators on Facebook last week. It’s worth noting that YouTube and TikTok also offer tools to creators to help them brainstorm ideas. For instance, YouTube Studio features an Inspiration tab that uses AI to help creators generate video ideas, while TikTok offers creators an AI assistant that can brainstorm ideas and uncover trends.

The desktop version of Edits will give creators more precise control over the editing process as well as the ability to work on a larger screen, which can be helpful during more advanced editing workflows. The company says creators will be able to sync their workflows seamlessly between mobile and desktop devices.

The upcoming desktop version will also allow Edits to better compete with CapCut, which already offers a desktop version.

Among the new features launching today is a Beta tab, which will provide creators with early access to experimental features that are still in development and allow them to provide Meta with feedback. The rollout of the Beta tab indicates that Meta wants to better compete with CapCut and accelerate feature development based on what creators actually want and will use.

Creators will also now be able to see more detailed metrics like their audience demographic breakdown and the time of day their audience is the most engaged. The new metrics join the app’s existing analytics, which include data such as how long viewers watch a video, how many followers were gained from a specific video, where users stop watching a certain video, and more.

Additionally, creators can search specific topics within the app’s Inspiration feed to discover reels and templates other creators are making around a given trend or idea. They’ll also be able to create multiple versions of a single piece of content to test what performs best before publishing.

Although Instagram didn’t share specific numbers about how many users Edits has, the company says that content made with the app sees a 10% higher save rate and 2% higher reshare rate compared to content not made on Edits, and that more than half of people watching reels on Instagram are seeing Edits-created content every day. 

Edits is free to download on iOS and Android.

The AI assistant announced today is currently in testing with attendees of Thursday’s creator event, while the desktop version of Edits is “coming soon,” Meta says. The rest of the features are launching to everyone today.

[This is a sponsored article with Synology.]

The cloud was supposed to eliminate infrastructure headaches. Instead, some businesses are discovering a new one: invoices they no longer understand.

Storage fees, data retrieval charges, and backup costs are quietly pushing cloud spending higher than many organisations anticipated—and artificial intelligence (AI) is about to make it significantly worse. 

It’s a challenge Taiwanese storage company Synology has been watching closely. 

At Computex 2026 in Taipei last week, the company argued that the economics of cloud-first infrastructure are beginning to shift.

The cloud bill that keeps growing

Businesses that moved enthusiastically to cloud-first infrastructure in the early 2020s are now sitting with bills that look nothing like the ones they signed up for. As costs continue to climb, some are reassessing whether keeping more data on-premise could make better financial sense.

The main catalyst is artificial intelligence. 

AI doesn’t just store data—it constantly accesses, moves, and processes it. Every inference call, every model training run, every search query is pulling data in and out of storage.

Gartner forecasts that more than 80% of enterprises will have deployed AI-enabled applications by 2026. And in a cloud environment, every one of those activities comes with a price tag.

From the start, cloud storage has been largely marketed on a simple figure: storage cost per gigabyte. Amazon Web Services S3 Standard, for example, runs at roughly US$23 per terabyte per month, with Google Cloud and Microsoft Azure in a similar range. 

For many businesses, the math looked straightforward. But what’s less visible is everything layered on top of that base rate.

Every time data is retrieved or moved—something AI workloads do constantly—cloud providers charge additional fees. 

On AWS, data egress starts from around US$0.09 per gigabyte transferred out. At scale, even restoring a 10TB dataset can quietly add about US$700 to the bill, before factoring in anything else.

Add API requests, cross-region replication, and backup-related charges, the total cost can be pushed up to four times the advertised storage rate. 

A Backblaze survey of more than 400 IT leaders found that 95% encountered unexpected cloud storage costs. And according to the Wasabi Global Cloud Storage Index, which surveyed 1,600 IT decision-makers globally, including 525 across APAC, 63% of organisations in the region exceeded their cloud storage budget in 2024. 

Businesses are bringing data back in-house

Amid rising costs, cloud repatriation—bringing data and workloads back from public cloud providers onto private or on-premise infrastructure—has moved from a niche IT discussion to a mainstream business decision. 

A 2025 Barclays CIO survey found that 86% of enterprise CIOs planned to shift at least some workloads back to private or on-premise systems, the highest level recorded. The Flexera 2025 State of the Cloud report similarly shows that 21% of workloads have already been repatriated, even as overall cloud spending continues to grow.

However, it’s important to note that most organisations aren’t abandoning cloud wholesale. 

Instead, they are taking a hybrid approach by keeping cloud platforms for global accessibility and collaboration, while bringing back workloads that involve heavy storage, protection, and processing. Backup and production storage are among the most commonly repatriated, as these are the areas where costs scale most quickly.

For Singapore businesses, there’s another factor at play: regulation. Under MAS Technology Risk Management guidelines and PDPA obligations, companies are expected to know where their data is stored and how it is being accessed.

That becomes harder in large public cloud setups, where data can be spread across multiple regions and servers. On-premise systems, by contrast, make it easier to keep track of exactly where information sits and who has access to it, since everything is managed within a company’s own infrastructure.

What Synology is bringing to the table

Synology is one of the companies building for this shift. 

The firm is best known for its Network Attached Storage (NAS) hardware—physical devices that store data locally while still functioning as a private cloud. 

At Computex 2026, it outlined how it is expanding its NAS ecosystem beyond storage into AI-enabled data management and backup infrastructure.

At the centre of this push is Synology’s next-generation DiskStation Manager (DSM), the operating system that powers every Synology NAS device.

The Taiwanese firm has spent more than two decades building NAS hardware and software. Today, it has shipped over 14 million systems worldwide, managing more than 400 exabytes of data. 

AI that stays in-house

At Computex 2026, the company announced the roadmap for the next generation of DiskStation Manager, DSM Agent 2.0, expanding it from a storage operating system into an intelligent data platform for governed, on-premises AI workflows. The goal is to turn DSM from a storage system into a smarter data platform that can support AI tools running on a company’s own infrastructure.

Instead of sending data to external cloud services, businesses can use their own data, such as files, system logs, and usage data, to power AI tools internally, while keeping everything under their control.

“The next generation of DSM leverages over two decades of expertise to create an AI-ready platform that keeps organisations firmly in control of their data,” said Philip Wong, Chairman and CEO of Synology.

Some AI features available include a conversational assistant for troubleshooting and system management. More advanced AI agents are also in development, designed to handle tasks such as email drafting, formula searches, meeting transcription, and real-time translation, although no release date has been announced yet.

As these capabilities expand, privacy becomes even more important in the age of AI. The system already includes a feature that masks sensitive data such as names, ID numbers, email addresses, and financial information locally before anything is sent to external AI providers like OpenAI or Azure AI. 

Future updates will go further, with support for fully on-premise large language models, where no data needs to leave the organisation’s infrastructure.

Synology’s infrastructure is already at work in Singapore

The value of on-premise data infrastructure is already clear for Singapore businesses using Synology. 

Food chain I Love Taimei, which has 17 outlets in Singapore, uses Synology’s DSM system to manage surveillance footage across all locations. This cuts management time by 65% and also allows the company to run AI-powered customer analysis without sending footage to the cloud.

LASALLE College of the Arts also uses Synology NAS for file storage and 4K video collaboration, allowing students and staff to access large project files easily across Mac computers without compatibility issues or rising costs.

Together, these examples show why some organisations are rethinking the assumption that everything belongs in the cloud.

Cutting backup costs without the cloud

The same push toward more controlled, on-premise infrastructure also extends to backup. At Computex, Synology introduced ActiveProtect Manager 2.0, a centralised backup system that will launch in Q3 2026.

The key issue it addresses is cost. Most backup services charge per server, virtual machine, or device. ActiveProtect instead charges for the hardware, with no extra per-workload fees.

In some cases, customers have seen a lower total cost of ownership. For example, Taiwanese media company Info Times reduced setup costs by 65% and cut storage needs by 75%. Toyota also reduced its backup data by 75% through better storage efficiency.

ActiveProtect 2.0 works with existing systems, so companies don’t need to replace their current setup. It also uses machine learning to detect unusual backup activity and help prevent ransomware infections from being restored. 

And because everything is stored locally, recovery is faster—taking hours instead of days—and there are no cloud data transfer fees.

The bigger picture

Cloud still has an important role to play, whether for global access, extra computing capacity, or supporting teams across different regions.

What’s changing is that businesses are becoming more selective about what they keep in the cloud. Rather than moving everything to a single platform, many are deciding where data should live based on cost, performance, and compliance requirements.

For Singapore businesses that have quietly accepted rising cloud bills as part of the cost of doing business, it may be time to take a closer look at the numbers.

Explore Synology’s solutions here. 

Featured Image Credit: Synology

In context: Anthropic’s latest release is really a story about control, not just capability. The company is offering two versions of the same underlying model: Claude Mythos 5 for a small circle of trusted partners, and Claude Fable 5 for everyone else. The split reflects a core challenge Anthropic is still trying to solve – how to deploy an extremely capable system into the wild without simultaneously handing attackers a new class of offensive tools.

Mythos has already shown what it can do when it is not heavily restricted. Since April, when an earlier preview was sent to about 150 organizations under the banner of Project Glasswing, users have reported more than 10,000 critical security flaws in their own systems. Those same capabilities could also be used by attackers looking to break in, rather than to patch security holes.

For that reason, Mythos 5 is staying behind the glass for now. Anthropic is keeping it in the hands of a “small group of cyberdefenders and infrastructure providers,” along with select biology researchers, and is coordinating with US government agencies as part of the rollout. Access is effectively on a need-to-know basis, with the company signaling that a broader “trusted access program” will come later.

Fable 5 is where Anthropic is testing what a general-purpose release of Mythos-class technology looks like under constraint. Technically, it runs on the same underlying model as Mythos 5, but with hard limits built in. The system is designed to refuse or redirect a long list of requests related to cybersecurity, biology, and chemistry. When those guardrails trigger, the query is silently routed to an older model, Claude Opus 4.8, instead.

Anthropic has also wired Fable 5 to watch for distillation, where a user tries to harvest large volumes of answers to train a smaller model of their own. If the system thinks that is happening, those requests are also redirected to Opus 4.8. In other words, the company is not only trying to control what the model will talk about, but also what others can learn from it.

Anthropic has been wrestling with these decisions for months. Diane Penn, the company’s head of product management, told Wired that testing and feedback since the April preview have helped shape the current strategy, even though it is still far from perfect.

“We’re trying to make improvements in a way that’s beneficial, even if we don’t have the perfect [solution] for every use case to start,” she says. “Out of all the different approaches, this emerged as the most viable and the best one. We just ended up feeling like this was the best product choice for users to get the maximum value out of Fable 5.”

For now, the filters are tuned to err on the side of over-blocking. Penn has acknowledged that some harmless queries will be routed to the older model. Anthropic says it wants to refine its classifiers over time but argues that this level of caution is the only way to justify a wider release at this stage.

The stakes are higher because Fable and Mythos are not just chatbots that respond to prompts and stop. Anthropic says both can run “unattended” for longer stretches than previous Claude models, carrying out sequences of instructions without constant supervision.

That shift toward more agent-like behavior could substantially boost software engineering and other technical work, especially given Fable 5’s stronger code generation and visual capabilities. But it also raises obvious questions about what happens if those capabilities are misused.

Anthropic’s pricing reflects how powerful it believes these systems are compared with its other models. Fable 5 and Mythos 5 cost $10 per million input tokens and $50 per million output tokens, roughly double the company’s other public models but still cheaper than the earlier Mythos Preview. The higher price reflects both the performance gains and the sense that these models are still positioned as specialized systems, not yet just another SKU in a growing catalog.

Around Anthropic, competitors are moving in a similar direction. OpenAI has rolled out its own advanced cybersecurity model to a small circle of partners and convened a working group that echoes Project Glasswing. Both companies are preparing for potential IPOs and are under pressure to show investors they can ship cutting-edge technology without triggering backlash over safety concerns.

Even some of the people watching from the outside say the unease is justified. Canadian finance minister François-Philippe Champagne told the BBC that public concern around Mythos stemmed from “it’s the unknown, unknown.”

Anthropic co-founder Jack Clark has made a similar point from the inside, arguing that the industry has not yet figured out how to slow itself down. “You want the option to be able to take your foot off the gas and put your foot on the brake,” he said. “Right now, it’s like the AI industry has a gas pedal, but it doesn’t have a brake pedal.”

Agent skills have become an important part of real-world AI applications, providing a mechanism — a set of instructions saved in a folder of text-based markdown (.md) files, usually — for models to adapt to specific enterprise use cases and complex workflows.

However, optimizing these skills is a slow process and faulty process, as they cannot be trained in the same way as the parameters of the underlying AI model. Instead, users typically must update them manually by retyping the instructions in each file, playing a “guessing game” as to what changes might improve agentic AI performance and reduce errors.

SkillOpt, a new, open source (MIT Licensed) framework developed by Microsoft, does one better: it introduces an optimizer designed for agent skills, turning the agent’s skill .md document as a trainable object that evolves based on performance feedback.

It uses deep-learning-style optimization to make it possible for the AI to systematically explore modifications to the document and find the best combination of instructions. Most importantly, it accomplishes this procedural adaptation without making changes to the underlying model’s weights.

On various industry benchmarks, SkillOpt outperforms existing baselines, significantly boosting accuracy for models like GPT-5.5 and Qwen. The result is a set of compact, transferable skill artifacts that allow AI agents to adapt to new domains effortlessly.

The challenge of optimizing agent skills

Agent skills package procedural knowledge into natural-language specifications, including domain heuristics, tool-use policies, output constraints, and known failure modes. These skills provide an external interface for agents to adapt to complex enterprise workflows. In practice, agent skills are stored as text documents and inserted into the agent’s context before execution.

One of the key benefits of skills is that they customize the behavior of the underlying model without changing its weights. However, the skill document itself needs to be tweaked and optimized to get the best performance out of the agent.

While deep learning relies on strict mathematical controls for stability, human prompt engineering often relies on trial and error. When attempting to automatically update a skill document based on feedback, the lack of mathematical discipline makes text highly volatile.

Yifan Yang, Senior Research SDE at Microsoft Research Asia, told VentureBeat that the problem is not making changes, but ensuring those changes are mathematically sound.

“The breaking point isn’t whether a team can change a skill, it’s that they can’t guarantee the change is an improvement,” Yang said. “Three failure modes recur: no step-size control, so skills drift; no validation, so a fix that reads as reasonable gets written in and can quietly regress performance; and no negative memory, so the same failed edit keeps coming back.”

To illustrate how easily performance can drop when edits aren’t mathematically validated, Yang noted that “an ungated rewrite pushed GPT-5.5 on SpreadsheetBench from 41.8 down to 41.1.”

According to Yang, these failure modes are amplified in multi-step workflows “because that’s where frontier models are weakest zero-shot. Not on reasoning, but on procedural discipline: format, self-verification, tool policy.”

Before SkillOpt, agent skills were primarily hand-crafted, generated in a single shot, or evolved through loosely controlled self-revision pipelines that could not reliably improve under feedback.

Prompt optimization methods like TextGrad and GEPA treat language artifacts as optimizable objects and use trajectory feedback to evolve prompts, but they focus on single-prompt configurations rather than generating persistent, reusable skill artifacts.

Meanwhile, skill evolution and discovery methods like EvoSkill and Trace2Skill convert agent execution experiences into trajectory lessons to refine skill folders, build domain-specific libraries, or perform evolutionary search.

None of them apply deep-learning-style controls, such as learning rates, validation gates, and momentum, which are necessary to continuously train a single, compact skill document.

Importing mathematical discipline to text

SkillOpt optimizes a text document through an iterative propose-and-test loop that separates the model executing the tasks from the model optimizing the skill. The process unfolds in several steps:

• SkillOpt starts with an initial skill document and a frozen target model (or harness), where the target model runs a batch of tasks to generate execution trajectories that act as the evidence for the current step.

• An offline optimizer model analyzes these trajectories, separating successes from failures into minibatches. Looking at a minibatch helps the model identify systematic procedural errors rather than one-off anomalies. Based on these patterns, the optimizer proposes structural add, delete, or replace edits to the skill document.

• The proposed edits are reviewed to filter out duplicates or contradictions, and the optimizer then ranks these candidate edits by their expected utility.

• Rather than applying all proposed changes, SkillOpt clips the list to a maximum edit budget for that step, generating a candidate skill.

• The candidate skill is evaluated on a held-out validation set using the target model. If the candidate improves the validation score, it is accepted and becomes the new current skill. If it fails, the edits are rejected and sent to a rejected-edit buffer, providing negative feedback so the optimizer knows not to repeat that mistake.

SkillOpt directly addresses the problem of treating text as a trainable object by importing mathematical concepts from deep learning. The creators note that “the deep-learning analogy is operational rather than decorative,” helping the framework avoid the instability issues associated with other optimization techniques.

The edit budget acts as a learning rate. By limiting how many edits can be applied at once, the skill version is prevented from moving too far from its previous state, preserving continuity while allowing new procedures to be acquired. 

Just like checking validation loss in deep learning, the strict held-out examples ensure that plausible-sounding text edits are only kept if they mathematically improve the agent’s actual performance on the validation split.

At the end of an epoch, SkillOpt performs a slow update by comparing tasks under the previous and current epoch’s skills. This acts like a momentum term, carrying durable, long-horizon procedural lessons forward while isolating them from the fast, step-level edits.

SkillOpt in action

To evaluate the technique in practice, researchers tested SkillOpt across different models, ranging from large-scale frontier models like GPT-5.5 to smaller closed and open models including GPT-5.4-mini and Qwen3.5-4B. They also deployed the skills within different execution harnesses, using plain chat as well as complex coding harnesses like the Codex CLI and Claude Code.

The evaluation spanned diverse industry benchmarks including single-round question-answering, multi-round code generation involving tool use, and multimodal document reasoning. SkillOpt was measured against multiple baselines ranging from a default no-skill setting to human-written skills and one-shot LLM-generated skills. It was also compared against advanced prompt-optimization and skill-evolution methods, specifically Trace2Skill, TextGrad, GEPA, and EvoSkill.

SkillOpt dominated across the board, proving highly effective on all 52 evaluated combinations of model, benchmark, and harness. It was particularly effective with frontier models, delivering an average absolute improvement of +23.5 points against the no-skill baseline on GPT-5.5. Furthermore, SkillOpt outperformed a hypothetical oracle baseline that cherry-picks the best competing method for every problem.

Small target models saw immense relative gains, proving that a compact text file can supply procedural knowledge that small models lack in their weights. For example, GPT-5.4-nano nearly doubled its score on multimodal document QA and tripled its score on embodied interaction and sequential decision-making.

These academic benchmarks map to critical enterprise pain points. Zero-shot models often hallucinate formatting or fail to use tools properly in multi-step scenarios. Yang explained that the biggest performance leaps occurred in operations that enterprises historically struggle to automate reliably.

“Document data extraction… exact figures out of contracts, invoices, and forms — AP automation, claims, compliance,” Yang said. “What improves is reliability: precise formatting, self-verification, auditable outputs. And the gains come from learning procedure, not memorizing answers.”

For enterprise practitioners, the true value of SkillOpt lies in its portability, efficiency, and compatibility with existing infrastructure. Experiments confirm that the framework is harness-agnostic. In addition to basic chat, the same optimization loop was successfully integrated into tool-backed execution environments like the Codex CLI and Claude Code with significant gains on industry benchmarks.

Developers can train a skill using one execution loop and deploy it in another. For example, a spreadsheet skill trained entirely inside the Codex loop was moved directly into Claude Code and drove a +59.7 point gain over Claude Code’s native baseline without any further changes.

SkillOpt artifacts also transfer cleanly across model scales. A skill optimized for GPT-5.4 was deployed onto the smaller GPT-5.4-mini and GPT-5.4-nano models with positive gains, proving that the learned procedures encode reusable workflows rather than just exploiting quirks of a specific model’s architecture.

Finally, the framework is highly efficient regarding token usage and context window real estate. Across all benchmarks, the final deployed skills never exceeded 2,000 tokens, with a median length of roughly 920 tokens. This results in highly readable, auditable artifacts that a human practitioner can review and manage in minutes.

Implementation strategies and the enterprise ‘catch’

For enterprise tech leaders, adopting a new framework requires understanding the overhead and limitations. While the research paper notes that training tokens can reach up to 210 million for academic benchmarks, the reality for day-to-day enterprise use cases is much lighter. The high token counts in testing were largely due to re-scoring massive held-out test sets.

“The real upfront work is the verifier and a representative held-out split. The optimizer is light; the evaluation harness is where the engineering goes,” Yang said. He added that for everyday use, “in community frameworks like GBrain, where SkillOpt updates run on Claude Sonnet, training a skill for a single task averages just $1–5.” This optimization cost is a one-time fee that amortizes completely at deployment.

However, the framework requires specific conditions to work effectively, namely a few dozen representative examples and a scorable feedback signal. Teams should avoid applying SkillOpt to open-ended or subjective tasks. “With no clean automatic scorer you have to design a human- or model-based evaluator and watch its stability,” Yang said.

SkillOpt also integrates smoothly with existing orchestration stacks, removing a major adoption hurdle. For instance, developers already using pipeline compilers can run both systems harmoniously. “DSPy is a different, complementary layer,” Yang said. “It compiles declarative LM pipelines and optimizes program structure; SkillOpt optimizes the external skill state a frozen agent loads. You can run them together.”

Looking ahead, open-source developers are already scheduling SkillOpt to run periodically over their agents’ past trajectories, creating a small ecosystem of self-optimizing code-agent plugins. This continuous feedback loop represents a significant shift in how AI systems adapt.

“The valuable version of self-improvement is an agent autonomously discovering knowledge to improve its own behavior and the user experience, under verification and audit,” Yang said. “Skills are the fastest, cheapest, most reversible first step, and the same mindset points toward agents eventually optimizing themselves, all the way down to their own weights.”