In a landmark case, a jury found this week that Meta and YouTube negligently designed their platforms and harmed the plaintiff, a 20-year-old woman referred to as Kaley G.M. The jury agreed with the plaintiff that social media is addictive and harmful and was deliberately designed to be that way. This finding aligns with my view as a clinical psychologist: that social media addiction is not a failure of users, but a feature of the platforms themselves. I believe that accountability must extend beyond individuals to the systems and incentives that shape their behavior.

In my clinical practice, I regularly see patients struggling with compulsive social media use. Many describe a pattern of “doomscrolling,” often using social media to numb themselves after a long day. Afterwards, they feel guilty and stressed about the time lost yet have had limited success changing this pattern on their own.

It’s easy to understand why scrolling can be so addictive. Social media interfaces are built around a powerful behavioral mechanism known as intermittent reinforcement, says Judson Brewer, an addiction researcher at Brown University, which is the strongest and most effective type of reinforcement learning. This is the same mechanism that slot machines rely on: Users never know when the next reward—a shower of quarters, or a slew of likes and comments—will appear. Not all the videos in our feeds captivate us, but if we scroll long enough, we are bound to arrive at one that does. The ongoing search for rewards ensnares us and reinforces itself.

Why Social Media Feels Addictive

Individuals typically struggle on their own to address compulsive social media use. This should be no surprise, as habits are not typically broken through sheer discipline but rather by altering the reinforcement loops that sustain them. Brewer argues that “there’s actually no neuroscientific evidence for the presence of willpower.” Placing the burden to self-regulate solely on users misses the deeper issue: These platforms are engineered to override individual control.

A growing body of research identifies social media use and constant digital connectivity as important influences on the growing incidence of adolescent mental health problems. Brewer notes that adolescents are particularly vulnerable, as they are in a “developmental phase” in which reinforcement learning processes are especially strong. This vulnerability can be exploited by the design features of large social media platforms.

How Platforms Are Designed to Maximize Engagement

NPR uncovered records from a recent lawsuit filed by Kentucky’s attorney general against TikTok. According to these documents, TikTok implemented interface mechanisms such as autoplay, infinite scrolling, and a highly personalized recommendation algorithm that were systematically optimized to maximize user engagement.

TikTok’s algorithmically tailored “For You” content continuously tracks user behaviors, such as how long a video is watched, whether it is replayed, or quickly skipped. The feed then curates short videos, or reels, for the user based on past scrolling behavior and what is most likely to hold attention.

These documents show one example of a tech company knowingly designing products to maximize attention. I believe social media companies also have the capacity to reduce addictiveness through intentional design choices.

How Governments Are Regulating Social Media

The good news is we are not helpless. There are multiple levers for change: how we collectively talk about social media, how our governments regulate its design and access, and how we hold companies accountable for practices that shape user behavior.

Some countries are moving quickly to set policy around social media use. Australia has imposed a minimum age of 16 for social media accounts, with similar bans pending in Denmark, France, and Malaysia.

These bans typically rely on age verification. Users without verified accounts can still passively watch videos on platforms like YouTube, but this approach removes many of the most addictive features, including infinite scroll, personalized feeds, notifications, and systems for followers and likes. At the same time, age verification may cause different problems in the online ecosystem.

Other countries are targeting social media use in specific contexts. South Korea, for example, banned smartphone use in classrooms. And the United Kingdom is taking a different approach; its Age Appropriate Design Code instructs platforms to prioritize children’s safety while designing products. The code includes strong privacy defaults, limits on data collection, and constraints on features that nudge users toward greater engagement.

How Social Media Platforms Could Be Redesigned

A report called Breaking the Algorithm, from Mental Health America, argues that social media platforms should shift from maximizing engagement to supporting well-being. It calls for revamping recommendation systems to spot patterns of unhealthy use and adjusting feeds accordingly—for example, by limiting extreme or distressing content.

The report also argues that users should not have to intentionally opt out of harmful design features. Instead, the safest settings should be the default. The report supports regulatory measures aimed at limiting features such as autoplay and infinite scroll while enforcing privacy and safety settings.

Platforms could also give users more control by adding natural speed bumps, such as stopping points or break reminders during scrolling. Research shows that interrupting infinite scroll with prompts such as “Do you want to keep going?” substantially reduces mindless scrolling and improves memory of content.

Some social media platforms are already experimenting with more ethical engagement. Mastodon, an open-source, decentralized platform, displays posts chronologically rather than ranking them for engagement, and does not offer algorithmically generated feeds like “For You.” Bluesky gives users control by letting them customize their own algorithms and toggle between different feed types, such as chronological or topic-based filters.

In light of the recent verdict, it is time for a national conversation about accountability for social media companies. Individual responsibility will always be important, but so are the mechanisms employed by big tech to shape user behavior. If social media platforms are currently designed to capture attention, they can also be designed to give some of it back.

The cybersecurity industry needs to ‘fight fire with fire’ when it comes to AI, according to Integrity360’s Brian Martin.

Earlier this month, cybersecurity company Integrity360 held the Dublin edition of its global cyber conference, Security First.

Held on 10 March, the conference saw cybersecurity experts and companies from across the country flock to the Aviva Stadium to sit in on a number of keynotes and panels, all exploring a wide variety of pressing cyber topics and trends.

Richard Ford, CTO at Integrity360, told SiliconRepublic.com at the Aviva that the Security First series of events provides an opportunity to bring some thought leaders from across the cyber industry together.

“From inside Integrity as well, but also from across industry because I think it’s important to get multiple voices,” he said.

But what does ‘security first’ mean?

Ford told us that the term refers to Integrity360’s model and view of cybersecurity, which is “to enable organisations to look at the full life cycle of cybersecurity, and look at it holistically, rather than just looking at one single segment”.

He explained that cybersecurity needs to be a whole-of-organisation priority, rather than just an IT concern.

“You know, it can’t just be a technology, it can’t be the technologists, the analysts, the engineers. It has to come from the very top of the organisation to think, actually, cybersecurity is important, it’s one of our key risks to our business,” he said.

“So we need to take that seriously and put the right effort, the right spend and the right focus on it to make sure that we are not the next organisation to be in the headlines and get breached.”

Human-AI era

This year, the conference theme was ‘Resilience Redefined: Securing the Human-AI Era’, with the event examining “how AI, machine identities and changing regulations are reshaping organisational defence”, according to the company.

While AI has certainly become a top opportunity for today’s cybersecurity teams, it has also introduced considerable cyberthreats as well.

“It’s not so much that AI is actually finding new things to exploit, but rather that it can automate a lot of the activities that normally would be performed by human attackers,” explained Brian Martin, Integrity360’s director of product management.

“So therefore, they can find exposures a lot faster, they can adapt more quickly to blockers that they find, and then they can execute more and more stages of the attack autonomously and at scale.”

Chris Hosking, AI and cloud security evangelist at SentinelOne, said AI has impacted cybersecurity by “empowering the threat landscape”.

“It’s made [threat actors] faster, it’s given them more sophistication than they’ve had before and it’s also lowered the barrier to entry,” he said.

However, he emphasised that AI also provides considerable opportunities for cyber defence teams. “We can have a phenomenal edge if we move to things like agentic auto triage, auto investigation. But we’re really sort of only dipping our toes right now into what AI can do for us, but a lot of opportunity ahead.”

Martin agreed, stating that cybersecurity teams need to “fight fire with fire” by utilising AI technology.

“We need to be able to move away from some of the manual and sort of fragmented visibility that we’ve had in order to be able to respond to that increasing scale and rapidity of attack.”

With the scale and complexity of cybersecurity and cyberthreats greater than it’s ever been, plenty in attendance at the conference had advice for companies that might be wondering how to adapt.

Niall Errity, director of professional services at Vectra, said that companies need to “understand their own gaps in their own network”.

“I think a lot of organisations tend to be very one or two tool-focused,” he said. “They think, ‘oh, I have an EDR [endpoint detection and response] installed on my endpoints and I’m perfectly fine now, I don’t need to worry about anything else beyond that’.

“Unless they’re doing the proper analysis in their own environments, doing proper testing to really make sure they’re resilient, they won’t really know their own gaps. So my recommendation would be to constantly look at what’s happening in the news, what’s happening with organisations. Take those learnings and test your own environments to make sure that you have those gaps closed.”

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.

Embedding fasteners or other hardware into 3D prints is a useful technique, but it can bring challenges when applied to large or non-flat objects. The solution? Use a gap-cap.

The gap-cap technique is essentially a 3D printed lid. One pauses a print, inserts hardware, then covers it with a lid before resuming the print. The lid — or gap-cap — does three things. It seals in the part, it fills in empty space left above the component, and it provides a nice flat surface for subsequent layers which makes the whole process much cleaner and more reliable.

This whole technique is a bit reminiscent of the idea of manual supports, except that the inserted piece is intended to be sealed into the print along with the embedded hardware under it.

If you have never inserted anything larger than a nut or small magnet into a 3D print, you may wonder why one needs to bother with a gap-cap at all. The short version is that what works for printing over small bits doesn’t reliably carry over to big, odd-shaped bits.

For one thing, filament generally doesn’t like to stick to embedded hardware. As the size of the inserted object increases, especially if it isn’t flat, it increasingly complicates the printer’s ability to seal it in cleanly. Because most nuts are small, even if the printer gets a little messy it probably doesn’t matter much. But what works for small nuts won’t work for something like an LED strip mounted on its side, as shown here.

In cases like these a gap-cap is ideal. By pre-printing a form-fitting cap that covers the inserted hardware, one provides a smooth and flat surface that both seals the component in snugly while providing an ideal surface upon which to resume printing.

If needed, a bit of glue can help ensure a gap-cap doesn’t shift and cause trouble when printing resumes, but we can’t help but recall the pause-and-attach technique of embedding printed elements with the help of a LEGO-like connection. Perhaps a gap-cap designed in such a way would avoid needing any kind of adhesive at all.

Processing 200,000 tokens through a large language model is expensive and slow: the longer the context, the faster the costs spiral. Researchers at Tsinghua University and Z.ai have built a technique called IndexCache that cuts up to 75% of the redundant computation in sparse attention models, delivering up to 1.82x faster time-to-first-token and 1.48x faster generation throughput at that context length.

The technique applies to models using the DeepSeek Sparse Attention architecture, including the latest DeepSeek and GLM families. It can help enterprises provide faster user experiences for production-scale, long-context models, a capability already proven in preliminary tests on the 744-billion-parameter GLM-5 model.

The DSA bottleneck

Large language models rely on the self-attention mechanism, a process where the model computes the relationship between every token in its context and all the preceding ones to predict the next token.

However, self-attention has a severe limitation. Its computational complexity scales quadratically with sequence length. For applications requiring extended context windows (e.g., large document processing, multi-step agentic workflows, or long chain-of-thought reasoning), this quadratic scaling leads to sluggish inference speeds and significant compute and memory costs.

Sparse attention offers a principled solution to this scaling problem. Instead of calculating the relationship between every token and all preceding ones, sparse attention optimizes the process by having each query select and attend to only the most relevant subset of tokens.

DeepSeek Sparse Attention (DSA) is a highly efficient implementation of this concept, first introduced in DeepSeek-V3.2. To determine which tokens matter most, DSA introduces a lightweight “lightning indexer module” at every layer of the model. This indexer scores all preceding tokens and selects a small batch for the main core attention mechanism to process. By doing this, DSA slashes the heavy core attention computation from quadratic to linear, dramatically speeding up the model while preserving output quality.

But the researchers identified a lingering flaw: the DSA indexer itself still operates at a quadratic complexity at every single layer. Even though the indexer is computationally cheaper than the main attention process, as context lengths grow, the time the model spends running these indexers skyrockets. This severely slows down the model, especially during the initial “prefill” stage where the prompt is first processed.

Caching attention with IndexCache

To solve the indexer bottleneck, the research team discovered a crucial characteristic of how DSA models process data. The subset of important tokens an indexer selects remains remarkably stable as data moves through consecutive transformer layers. Empirical tests on DSA models revealed that adjacent layers share between 70% and 100% of their selected tokens.

To capitalize on this cross-layer redundancy, the researchers developed IndexCache. The technique partitions the model’s layers into two categories. A small number of full (F) layers retain their indexers, actively scoring the tokens and choosing the most important ones to cache. The rest of the layers become shared (S), performing no indexing and reusing the cached indices from the nearest preceding F layer.

During inference, the model simply checks the layer type. If it reaches an F layer, it calculates and caches fresh indices. If it is an S layer, it skips the math and copies the cached data.

There is a wide range of optimization techniques that try to address the attention bottleneck by compressing the KV cache, where the computed attention values are stored. Instead of shrinking the memory footprint like standard KV cache compression, IndexCache attacks the compute bottleneck. 

“IndexCache is not a traditional KV cache compression or sharing technique,” Yushi Bai, co-author of the paper, told VentureBeat. “It eliminates this redundancy by reusing indices across layers, thereby reducing computation rather than just memory footprint. It is complementary to existing approaches and can be combined with them.”

The researchers developed two deployment approaches for IndexCache. (It is worth noting that IndexCache only applies to models that use the DSA architecture, such as the latest DeepSeek models and the latest family of GLM models.)

For developers working with off-the-shelf DSA models where retraining is unfeasible or too expensive, they created a training-free method relying on a “greedy layer selection” algorithm. By running a small calibration dataset through the model, this algorithm automatically determines the optimal placement of F and S layers without any weight updates. Empirical evidence shows that the greedy algorithm can safely remove 75% of the indexers while matching the downstream performance of the original model.

For teams pre-training or heavily fine-tuning their own foundation models, the researchers propose a training-aware version that optimizes the network parameters to natively support cross-layer sharing. This approach introduces a “multi-layer distillation loss” during training. It forces each retained indexer to learn how to select a consensus subset of tokens that will be highly relevant for all the subsequent layers it serves.

Real-world speedups on production models

To test the impact of IndexCache, the researchers applied it to the 30-billion-parameter GLM-4.7 Flash model and compared it against the standard baseline.

At a 200K context length, removing 75% of the indexers slashed the prefill latency from 19.5 seconds down to just 10.7 seconds, delivering a 1.82x speedup. The researchers note these speedups are expected to be even greater in longer contexts.

During the decoding phase, where the model generates its response, IndexCache boosted per-request throughput from 58 tokens per second to 86 tokens per second at the 200K context mark, yielding a 1.48x speedup. When the server’s memory is fully saturated with requests, total decode throughput jumped by up to 51%.

For enterprise teams, these efficiency gains translate directly into cost savings. “In terms of ROI, IndexCache provides consistent benefits across scenarios, but the gains are most noticeable in long-context workloads such as RAG, document analysis, and agentic pipelines,” Bai said. “In these cases, we observe at least an approximate 20% reduction in deployment cost and similar improvements in user-perceived latency.” He added that for very short-context tasks, the benefits hover around 5%.

Remarkably, these efficiency gains did not compromise reasoning capabilities. Using the training-free approach to eliminate 75% of indexers, the 30B model matched the original baseline’s average score on long-context benchmarks, scoring 49.9 against the original 50.2. On the highly complex AIME 2025 math reasoning benchmark, the optimized model actually outperformed the original baseline, scoring 92.6 compared to 91.0.

The team also ran preliminary experiments on the production-scale 744-billion-parameter GLM-5 model. They found that eliminating 75% of its indexers with the training-free method yielded at least a 1.3x speedup on contexts over 100K tokens. At the same time, the model maintained a nearly identical quality average on long-context tasks.

Getting IndexCache into production

For development teams wanting to implement the training-free approach today, the process is straightforward but requires careful setup. While the greedy search algorithm automatically finds the optimal layer configuration, the quality of that configuration depends on the data it processes.

“We recommend using domain-specific data as a calibration set so that the discovered layer-sharing pattern aligns with real workloads,” Bai said.

Once calibrated, the optimization is highly accessible for production environments. Open-source patches are already available on GitHub for major serving engines. “Integration is relatively straightforward — developers can apply the patch to existing inference stacks, such as vLLM or SGLang, and enable IndexCache with minimal configuration changes,” Bai said.

While IndexCache provides an immediate fix for today’s compute bottlenecks, its underlying philosophy points to a broader shift in how the AI industry will approach model design.

“Future foundation models will likely be architected with downstream inference constraints in mind from the beginning,” Bai concluded. “This means designs that are not only scalable in terms of model size, but also optimized for real-world throughput and latency, rather than treating these as post-hoc concerns.”