When it comes to Chinese AI labs using distillation techniques to extract knowledge from frontier model makers, Y Combinator CEO Garry Tan is hoping regulators stay out of it. In fact, he thinks U.S. AI labs should perhaps play the same game.

“I would do nothing,” he told CNBC in an interview earlier this week. “We could argue that there should be an American distillation regime.”

He elaborated to TechCrunch that this means he wants smaller, American open-weight AI labs to use the same kind of training techniques on American frontier AI labs, giving the U.S. a more robust set of open-weight options that aren’t Chinese.

Distillation is when a model maker extensively prompts another model in order to learn how it works and reasons. It is commonly, and legitimately, used by AI labs to help train new models.

Anthropic this week released its second report alleging that Chinese labs are engaged in “illicit distillation attacks,” hiding their identities to distill without permission and relying on fraud and stolen credentials to do so. Anthropic CEO Dario Amodei had previously publicly called on U.S. regulators to crack down on distillation.

It’s notable that the commander of Silicon Valley’s prestigious and prolific startup accelerator doesn’t agree.

To be clear, Tan isn’t advocating for American AI labs to use stolen credentials to distill. He wants them to be free to come in the front door. In fact, his argument is twofold. He feels it’s an overreach for AI labs to dictate what their customers can do with the information their models share with them.

He also notes that the proprietary AI labs didn’t ask permission when they vacuumed up as much human knowledge as they could to train their models. They famously ingested plenty of copyrighted material without the permission of those intellectual property holders.

“Controlling what users and customers do with API calls to closed weight models feels constraining, and there’s a role government can play here to normalize the fact that access to intelligence that was trained on broad public access data should itself also be more a form of a public good than something locked away behind restrictive terms of service,” he told TechCrunch when asked why American labs should be free to distill, too.

Tan, who is himself such an avid AI user that he once described himself as having cyber psychosis, wants to see a balance between open-weight AI labs and frontier labs.

“They are at the frontier and driving it forward. We want that to be fundable, and be a great business model ongoing,” he told CNBC. “You want open weight models to give people freedom and access.”

To him, the true AI doomer scenario is for all the immense power of frontier AI to wind up in the hands of a single powerful, proprietary provider. “The nightmare scenario, the doomer scenario for AI is that there’s just one company,” he said. “It has the best access to capital. It has the best AI researchers. It runs away with it and suddenly there’s one company that’s monolithic. And that would be bad.”

The studio produced some early hits for the PSVR and Oculus Quest.

A VPN normally adds some network overhead, but a severe slowdown is usually caused by server distance or congestion, protocol choice, weak Wi-Fi, packet loss, MTU problems, device limits, or a poor route between your internet provider and the VPN server. The fastest way to diagnose the problem is to compare the same connection with the VPN off and on, then change one variable at a time.

Do not begin by changing advanced settings. Establish a baseline first. If the underlying connection is already slow, changing VPN servers, protocols, or packet sizes will not fix the real bottleneck.

Quick Fixes to Try First

These checks solve many VPN speed problems without changing low-level network settings. Work through them in order and retest after each change.

The important troubleshooting rule is to change only one variable before each test. If you switch the server, protocol, Wi-Fi band, and DNS settings simultaneously, you will not know which change actually affected performance.

First Confirm the VPN Is Actually the Problem

A VPN cannot provide more usable capacity than the connection underneath it. VPN traffic still has to cross your Wi-Fi or Ethernet connection, your internet service provider, and the wider internet before reaching the VPN server.

VPNs also add processing and an additional network path. Cloudflare’s explanation of VPN speed factors notes that encryption and routing through an intermediary VPN server can increase latency and affect performance.

Run a Baseline Speed Test

Disconnect the VPN and test the connection on the same device and network you normally use. Record at least:

• Download speed: how quickly data reaches your device.
• Upload speed: how quickly your device sends data.
• Latency or ping: how long a request takes to travel to a destination and return.
• Packet loss: the proportion of packets that fail to reach their destination, if the testing tool reports it.

Our internet speed testing guide explains how to read download, upload, ping, and jitter results if you need a baseline before starting the VPN-specific diagnosis.

Reconnect the VPN and repeat the test under otherwise similar conditions. One result can be noisy, so several runs are more useful than one unusually good or bad measurement.

Read the Pattern, Not Just the Number

If both VPN-off and VPN-on results are poor, investigate the underlying internet connection first. If the connection is fast without the VPN but consistently much slower with it, the VPN path deserves closer inspection.

Different patterns point toward different causes. One slow VPN server suggests server congestion or routing. Every server being slow may indicate protocol, device, router, or provider limitations. A fast speed test combined with websites that stall halfway through loading can point toward packet loss, Domain Name System issues, or Maximum Transmission Unit problems rather than raw bandwidth.

10 Reasons Your VPN Is Slow and How to Fix Each One

1. The VPN Server Is Too Far Away

Distance usually affects VPN performance because your packets have to travel to the VPN server before continuing toward their final destination. The longer that round trip becomes, the more latency is added.

Latency is the delay between sending network traffic and receiving a response. It matters especially for gaming, video calls, remote desktops, and interactive web applications because these activities depend on frequent back-and-forth communication.

For example, if you are in Washington, D.C. and connect to a VPN endpoint in Tokyo, traffic must cross a long international path even when the website you are opening is hosted nearby in the United States.

Fix: connect to the nearest suitable city or country unless you specifically need a distant exit location. Proton’s current VPN speed guidance similarly recommends nearby servers and notes that distance and server load affect connection performance.

The geographically closest server is not guaranteed to be the fastest. Internet routing depends on peering arrangements and network topology, so test two or three nearby options rather than assuming the nearest city will always win.

2. The VPN Server Is Overloaded

A VPN server and its network connection have finite capacity. When many customers use the same endpoint at once, they compete for processing resources and available bandwidth.

The resulting symptom can be confusing because the VPN may work normally in the morning and become much slower during evening peak hours.

Fix: switch to another server in the same general region. If your VPN app displays server-load indicators, compare alternatives from that provider. Load percentages should only be treated as relative indicators within the provider because different companies calculate and expose them differently.

Suppose one London server is slow but another London endpoint performs normally on the same device and protocol. That strongly points toward server-specific congestion or routing rather than a problem with your home connection.

ExpressVPN’s current slow-speed troubleshooting similarly recommends moving to another VPN location when one endpoint performs poorly.

3. Your VPN Protocol Is a Poor Fit for the Network

A VPN protocol defines how your device establishes and transports the encrypted tunnel. Common options include WireGuard, OpenVPN, Internet Key Exchange version 2, usually called IKEv2, and provider-specific protocols.

Different protocols behave differently across operating systems, networks, firewalls, and connection types. A protocol that performs well on fiber may not behave the same way on a restrictive hotel network or unstable cellular link.

Fix: start with the VPN application’s automatic or recommended protocol. If performance remains poor, test the alternatives your provider supports and record the result of each one.

WireGuard is often a sensible option to test because it has a relatively compact design, but it should not be described as universally fastest. Implementation, hardware, server capacity, route quality, and the underlying network still matter.

OpenVPN can operate over User Datagram Protocol, or UDP, and Transmission Control Protocol, or TCP. UDP generally avoids some retransmission behavior of TCP and is commonly selected for performance-sensitive VPN use. TCP can be useful where network restrictions or reliability requirements favor it, but tunneling TCP traffic inside another TCP transport can behave poorly on lossy connections because multiple reliability mechanisms react to the same packet loss.

4. Your Wi-Fi Is the Real Bottleneck

A stable VPN cannot repair a weak wireless connection. Distance from the router, walls, radio interference, crowded channels, and a weak access point can all reduce speed or increase packet loss before VPN traffic even reaches the internet.

Fix: if possible, repeat the same VPN test over Ethernet. A large improvement over the wired connection identifies Wi-Fi as an important part of the problem.

You can also:

• move closer to the access point;
• try the 5 GHz or 6 GHz band when signal strength is sufficient;
• use 2.4 GHz when you need greater range and higher frequencies are weak;
• reduce local interference;
• test another trusted network.

A common diagnostic result is a laptop reaching high throughput over Ethernet with the VPN enabled but slowing dramatically from a distant bedroom over Wi-Fi. In that case, changing VPN providers is unlikely to fix the wireless bottleneck.

If the issue appears mainly on airports, hotels, or cafes, the existing public Wi-Fi safety guide covers additional network-selection and security checks.

5. Your Base Internet Connection Is Congested

VPN performance depends on the internet connection underneath it. A household connection that is already saturated by large downloads, cloud backups, video calls, or operating-system updates has less capacity available for the VPN.

Upload congestion deserves particular attention. When upstream traffic is fully saturated, acknowledgments and other small packets can be delayed, making downloads and interactive applications feel slower even when substantial downstream bandwidth remains available.

Fix: pause large background transfers and retest. Check other computers, consoles, televisions, phones, network-attached storage devices, and cloud-sync applications on the same connection.

Also compare performance at another time of day. If both VPN and non-VPN speeds collapse during the same evening period, the root cause is more likely to be local or ISP congestion than the VPN tunnel itself.

6. Packet Loss or Poor Routing Is Reducing Throughput

Packet loss means some network packets fail to reach their destination. Reliable protocols have to recover from those losses, which can reduce effective throughput and increase delay.

Common symptoms include:

• video calls breaking up;
• gaming latency jumping unpredictably;
• downloads repeatedly speeding up and slowing down;
• one VPN region performing much worse than another;
• remote desktop sessions freezing briefly.

The route between your internet provider and a particular VPN data center may also be inefficient even when normal non-VPN traffic follows a better route to the same general part of the internet.

Fix: test another VPN server, preferably in the same region first. Then compare another network, such as Ethernet versus Wi-Fi or a trusted mobile connection. If one server is consistently problematic while nearby alternatives are normal, report that server and the approximate test time to the VPN provider.

Do not conclude that an ISP is deliberately throttling VPN traffic simply because one route performs poorly. Congestion, peering problems, packet loss, and temporary routing changes can produce similar symptoms.

7. MTU or Fragmentation Problems Are Affecting Larger Packets

This cause is less visible but important when a VPN connects successfully and then certain websites, uploads, or larger transfers stall.

Maximum Transmission Unit, or MTU, is the largest packet size an interface or network path can carry under a particular configuration. A VPN adds its own headers and encapsulation overhead, which increases the size of the packet sent across the underlying network.

If a path cannot carry that resulting packet and the network’s Path MTU Discovery process fails, larger packets may be dropped or repeatedly fragmented.

OpenVPN’s current OpenVPN 2.7 manual explicitly describes the classic symptom: the VPN starts correctly but stalls during active use when Path MTU Discovery is broken. It documents controls such as mssfix and fragmentation for specific OpenVPN UDP configurations.

Beginner fix: do not start by entering arbitrary MTU values. First update the VPN software, reconnect, try another protocol, and use the provider’s automatic configuration. A protocol change can alter encapsulation behavior enough to avoid the problematic path.

Practitioner fix: if you administer the tunnel or router, investigate the actual path MTU and Maximum Segment Size rather than copying a value from an unrelated connection. Cisco’s documentation on IPsec pre-fragmentation likewise describes how tunneling overhead can cause downstream fragmentation around near-MTU packets.

OpenVPN-specific options should not be copied into WireGuard, IKEv2, or proprietary VPN configurations. Each tunnel technology handles packet sizing differently.

8. Your Device or Router Cannot Process the VPN Fast Enough

Encryption and tunnel processing must happen somewhere. If the VPN application runs on a laptop or phone, that device performs much of the work. If the VPN is configured directly on a router, the router may have to encrypt and route traffic for every device in the home.

This becomes noticeable on fast internet connections because a low-power router can hit its processing ceiling long before the internet plan reaches its advertised maximum.

Fix: compare the router-based VPN with the VPN application’s performance on a reasonably modern computer using the same general network and nearby server. If the laptop reaches much higher throughput, router processing may be the bottleneck.

Other clues include unusually high processor use, a router interface becoming sluggish during large transfers, or a very consistent VPN speed ceiling regardless of which fast server you select.

There is no universal throughput limit for VPN routers. Processor architecture, hardware acceleration, cipher implementation, tunnel protocol, firmware, packet size, and thermal limits all affect the result.

Some VPN features intentionally prioritize privacy, censorship resistance, or filtering rather than maximum throughput.

Examples include:

• multi-hop VPN routing;
• privacy-focused hardened entry servers;
• Tor routing;
• traffic obfuscation;
• advanced filtering or inspection features.

Multi-hop is the easiest example. Instead of passing through one VPN server, traffic passes through additional infrastructure before reaching the wider internet. That creates more distance and additional processing opportunities.

Fix: compare a normal single-hop connection with the advanced feature disabled only if your security requirements allow it. Do not automatically disable privacy or anti-censorship features simply to gain speed if they are important to your threat model.

If performance returns to normal with the extra feature off, the slowdown may be an expected operational trade-off rather than a fault.

10. The VPN App, Provider, or Free Tier Is the Bottleneck

After eliminating local network problems, the limiting factor may genuinely be the VPN service.

Possible causes include:

• an outdated application;
• a buggy VPN network adapter or driver;
• a temporary provider outage;
• congested data-center connectivity;
• a small free-server pool;
• free-tier speed restrictions;
• an account or plan that provides access to fewer high-capacity servers.

Fix: update the VPN app, restart it, test multiple nearby servers, compare supported protocols, and check the provider’s official service-status information if available.

Free services require particular care. Some impose a formal speed restriction. Others do not throttle bandwidth directly but expose fewer servers, which can leave free endpoints more heavily loaded.

If the same device and connection perform well without the VPN but remain dramatically slower across several VPN servers and protocols, the provider becomes a more plausible bottleneck.

Troubleshooting Table: Match the Symptom to the Likely Cause

The symptom often tells you where to test next. Use this table to avoid changing unrelated settings.

The useful pattern is comparative. A result that changes when you alter one server suggests a different cause from a result that stays identical across every VPN location.

How to Test Whether a Fix Worked

A good troubleshooting test changes one factor and holds the others as steady as reasonably possible.

1. Use the same device.
2. Stay on the same Wi-Fi or Ethernet network.
3. Use the same VPN location unless the server itself is the variable being tested.
4. Use the same speed-test service and nearby test destination.
5. Run several measurements instead of relying on one result.
6. Record download speed, upload speed, and latency.
7. Record packet loss or jitter if the tool provides them.

Suppose you are testing protocols. Leave the VPN server unchanged, switch from protocol A to protocol B, reconnect, and repeat the same measurements. If you are testing servers, leave the protocol unchanged and move only to another nearby server.

Do not rely on a universal rule such as “a VPN should lose no more than 10%.” The relative difference depends on your internet speed, server distance, device, protocol, network path, and workload. A 30 Mbps loss has very different implications on a 50 Mbps connection than on a 1 Gbps connection.

Latency may also matter more than headline download speed. A remote worker could prefer 200 Mbps with stable low latency over 400 Mbps with severe jitter if the main workload is video conferencing and interactive remote sessions.

When Slower VPN Speed Is Expected

Some performance loss is a normal consequence of the route or security features you deliberately selected.

• Distant VPN server: longer round trips increase latency.
• Multi-hop routing: traffic crosses additional VPN infrastructure.
• Tor-related routing: additional relay hops prioritize anonymity properties rather than raw speed.
• Obfuscation: extra processing may be necessary on networks trying to detect or block VPN traffic.
• Busy shared server: available capacity may vary with demand.
• Router-level tunneling: low-power hardware can limit encryption throughput.

Expected overhead should not become an excuse for accepting a severe unexplained slowdown. If a nearby single-hop server performs dramatically worse than the same internet connection without the VPN, controlled troubleshooting is still worthwhile.

Contact support after you have established a repeatable pattern rather than sending only “the VPN is slow.”

Useful diagnostic information includes:

• operating system and version;
• VPN application version;
• VPN protocol;
• server name, city, or country;
• approximate time of the problem;
• VPN-off and VPN-on speed results;
• whether the connection uses Wi-Fi or Ethernet;
• internet provider or network type;
• whether another nearby VPN server behaves normally.

If the slowdown began immediately after a software update, say so. If only one server is affected, identify that server. These details help separate application problems from server, routing, or local-network issues.

Do not send passwords, recovery codes, private keys, authentication tokens, or other credentials in diagnostic messages.

Key Takeaways

Frequently Asked Questions

September is typically the time when television gets serious again. The summer’s biggest TV shows have aired their finales, the fall schedule is kicking into high gear, and streamers are rolling out everything from high-concept sci-fi to prestige dramas. There’s plenty of familiar IP in the mix, but also some genuinely original shows worth putting on your to-watch list.

While plenty of outlets will tell you what everyone is watching (see: Reacher), that’s not our goal. As always, the TV shows we’re watching this month reflect the kinds of stories WIRED covers every day—shows that explore technology, science, politics, culture, and the sometimes unsettling ways they intersect.

From returning favorites to ambitious new series—and at least one show that might rekindle your love of traditional anime—these are the TV shows we think you should make time to stream this month.

The Paper

Amazon has made its Quick desktop assistant generally available on macOS and Windows, pitching it to enterprise IT as the governed alternative to shadow AI. Quick runs in ordinary European AWS regions with inference kept inside Europe, but it is not on the published service list for the sovereign cloud AWS opened in Brandenburg in January to put European data beyond the reach of US law.

Amazon has made its Quick desktop app generally available on macOS and Windows, AWS said, and added an activity feed to the iOS and Android versions that pools email, calendar, CRM and messaging into one list.

The pitch is governance. Amazon says customer data stays inside the customer’s own environment, conversations stay private, and audit trails run through CloudWatch and CloudTrail. It describes the underlying AWS as the infrastructure behind the world’s most security-sensitive workloads.

The compliance list is HIPAA, FedRAMP, SOC 2 and ISO 27001. The argument underneath is that shadow AI is what happens when staff route around approved tools, and that the answer is a sanctioned assistant rather than a ban.

In Europe that argument runs into a distinction Amazon drew itself in January.

Quick has run in the AWS Frankfurt region since March, with inference requests routed only within European regions. It also runs in London and Ireland.

Those are ordinary AWS regions.

The AWS European Sovereign Cloud is not. It opened in Brandenburg on 14 January, staffed only by EU residents and overseen by a board of EU citizens, on a commitment of EUR 7.8B.

Quick does not appear on its published service list. SageMaker and Bedrock do.

The distinction matters to the buyers Amazon is describing. Airbus is moving 70 applications off AWS to the French provider Scaleway, out of roughly 900 it runs, on a contract worth more than EUR 50M.

Its stated test was legal safeguards against non-European laws. The law in question is the US CLOUD Act, which reaches American companies’ data wherever in the world it is held.

Then there is the question of whether anyone pays. Of the Microsoft 365 users who can reach Copilot Chat, 3.3% pay for Copilot, about 15 million seats out of 450 million.

Amazon names three customers in the announcement, Southwest Airlines, Labcorp and the PGA Tour. All three quotes are supplied testimonials. The post gives no pricing, no seat numbers and no date for sovereign cloud support.

Quick’s case is that it is the assistant IT already controls. In Europe the follow-up is which AWS they control it from, and European rivals are selling the answer.

• CHUWI UniBox AI495 Pro carries up to 192GB of unified memory
• The Ryzen AI MAX+ PRO 495 reaches 5.2 GHz across 16 cores
• CHUWI’s new mini workstation can run 300 billion parameter models locally

CHUWI has unveiled the UniBox AI495 Pro at IFA Berlin 2026, combining AMD‘s new Ryzen AI MAX+ PRO 495 chip with unusually high memory capacity.

The compact machine can carry up to 192GB of LPDDR5x memory, matching the maximum memory offered by Apple‘s latest workstation.

BearingPoint’s Laura Kealy explores how she found herself working as a senior technology analyst.

Donegal native Laura Kealy found her secondary education abruptly interrupted in 2020 with the onset of the global pandemic. From there, she went on to pursue management science and information systems studies at Trinity College Dublin, before undertaking a graduate programme at management and tech consulting firm BearingPoint. 

“Maths and business were my favourite subjects in school, which led me to pursue a degree that combined both subjects alongside technology. While I enjoyed the technical side of my studies, I also knew I wanted a career that involved working closely with people and helping to solve real-world problems,” Kealy told SiliconRepublic.com. 

She explained that BearingPoint stood out to her after interacting with future team members at a college careers fair. Kealy said, “Everyone I spoke to was welcoming, genuine and passionate about what they did, which gave me a great impression of the company culture.”

Can you describe a typical day in your role? 

One of the things I enjoy most about my role is the variety. A typical day might involve meeting with clients to understand their requirements, collaborating with my teammates to solve technical challenges, and then carrying out hands-on work such as analysing data in Power BI, writing PowerShell scripts or creating project documentation.

I also enjoy the collaborative aspect of the work I do. Whether it’s working with teammates or engaging directly with clients, there’s always an opportunity to learn from others. I have been fortunate to work with people who are always willing to share their knowledge and help me develop my skills. 

Did your responsibilities and workload change as the programme progressed? 

At BearingPoint, I’ve always been encouraged to learn through hands-on experience and client-facing work. Being part of a small team meant I was trusted early on with responsibilities and given opportunities to make an impact, while still having the support of my colleagues around me.

As I’ve grown more confident in my role, my workload and responsibilities have naturally increased, progressing from supporting project activities to leading key client deliverables. I’ve also had the opportunity to collaborate with colleagues at all levels of the organisation, from fellow graduates right through to directors and partners.

Working with people from different backgrounds and experiences has taught me new ways of thinking and approaching challenges.

What skills have you developed since being part of the company’s programme?

My technical knowledge has grown significantly through hands-on project work, particularly in areas such as AI, data analysis and Microsoft technologies. I have had great opportunities to continue learning through certifications, training and knowledge-sharing from colleagues.

I have also developed my communication and consulting skills. I’ve become much more confident communicating with different audiences, whether that’s providing updates to my team, presenting to clients or explaining technical concepts to people from a non-technical background.

Learning how to adapt my communication style to different audiences has been one of the most valuable skills I’ve gained throughout the programme and one that I use every day.

What have been the most challenging parts of working life since taking up this programme?

One of the most challenging aspects of my role has been keeping up with the pace of change in technology, particularly with the rapid evolution of AI. Working in this space means continuously learning, especially when you’re helping clients understand how these technologies can be used effectively within their organisations.

What started as a challenge has become something I now really enjoy. Whether it is writing a simple prompt to Copilot or building an advanced AI agent that can complete tasks on your behalf, AI is constantly evolving and there’s always something new to learn and experiment with.

It’s also been rewarding to share what I’ve learned with others and help build confidence in using these tools across different teams and projects.

Would you recommend the graduate programme at this company to others? 

I would definitely recommend the graduate programme at BearingPoint. The combination of challenging work, learning opportunities and a genuinely supportive culture has made it a great place to start my career.

I have had the opportunity to work on large-scale projects, learn from experienced colleagues and develop both technically and professionally. I really enjoy the strong social aspect of the office, with everything from sports and social events to charity initiatives.

Those experiences have made it easy to build relationships and have been a really important part of my time here. 

What advice would you like to give to future graduates who are just starting out?

My advice would be to say yes to opportunities, even if they’re outside your comfort zone. This applies to projectwork, but some of the experiences I’ve learned the most from weren’t necessarily part of my day-to-day role.

Getting involved in different teams, initiatives and events has really helped me develop new skills, build confidence and form great connections across the organisation.

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.

For this year’s IEEE annual election, the IEEE Tellers Committee will recognize the Sections with the highest voter turnout (based on total eligible voters) in each Region with an incentive reward after the election results are accepted by the IEEE Board of Directors.The incentive initiative is managed by the Tellers Committee, which retains full authority over all rules, operations, and decisions regarding the program.

Incentive guidelines

The Section sizes and their respective reward amounts are:

• Large sections consist of more than 1,501 eligible voting members. The top-performing large section in each IEEE region will be rewarded with US $1,000.
  
• Medium sections consist of 501 to 1,500 eligible voting members. Each region’s top-performing medium section will receive $600.
  
• Small sections consist of 500 or fewer eligible voting members. The top-performing small section in each region will receive $400.

To learn more about the incentive program, visit the IEEE annual election website. If you haven’t voted in the 2026 elections, you can learn about the candidates and vote here. Send questions to: elections@ieee.org.