Myself and several others have experienced consistently terrible customer service from AI companies. Getting in touch with a human support agent is often impossible, leaving customers stuck in endless loops with AI support bots that repeatedly acknowledge issues and request more information without ever resolving anything.

When issues are raised, companies often refuse refunds and offer little to no meaningful resolution, even when their own technical bugs cause usage limits to malfunction. Rather than taking responsibility, they shift the blame onto customers and hide behind their terms and conditions to justify denying refunds.

This is unfair. Customers shouldn't have to pay for technical errors they didn't cause or waste time chasing support instead of actual solutions. AI companies need to take accountability, treat customers fairly, and resolve problems rather than simply acknowledge them. Terms and conditions should not be an excuse to evade responsibility.

A few companies I've seen this issue with include Anthropic, OpenRouter, and Hugging Face.

What companies have any of use faced this issue with?

It will bring together eight leading institutions and aims to create a network of more than 1,000 scientists.

Real-time translation could help people communicate across languages in work, travel, and everyday life.

But some conversations are meaningful partly because they are slow and imperfect. People use gestures, pauses, and misunderstandings to find out how the other person thinks.

If translation removes all friction, it may improve communication while also removing some of the patience that communication requires.

I’m very much in favor of translation tools. I’m just curious whether perfect convenience changes the relationships built through them.

Has instant translation made you more willing to talk to people from different language backgrounds?

All the big AI safety conversations are about existential risk or model capabilities. Meanwhile the data protection side is just... quiet. Training data scraped without consent, chat logs stored forever, personal info memorized by models. What would it even look like to do data protection properly here? Not the checkbox consent stuff, the real thing.

TL;DR: He had the "we're fucked" call with Matt, then checked how the data is actually captioned — and it doesn't hold up.

He built owned dataset with LiDAR, Unreal and volume stages, no peers' work trained on.

Do you remember that introductory banter between Q and James Bond – played by Ben Whishaw, and Daniel Craig, respectively, in Skyfall, 2012 – at the National Gallery, London?

Q: Well, I'll hazard I can do more damage on my laptop sitting in my pajamas before my first cup of Earl Grey than you can do in a year in the field.

Bond: Oh, so why do you need me?

Q: Every now and then a trigger has to be pulled.

Bond: Or not pulled. It's hard to know which in your pajamas.

That’s the last layer of judgement, right there.

In another scene back at the MI6 HQ, Whitehall in Spectre (2015), Q injected the “Smart Blood” nano-machines into 007’s blood stream for tracking, after Bond's unauthorized Mexico mission.

Q: Well it's lovely to see you, 007... the Smart Blood program is obviously still in its developmental phase. So we may experience the odd drop in coverage during the first 24 hours...

[Bond coldly glares at Q]

Q: …48 hours after administration, but after that it should work perfectly.

Later…

Bond: Could you do one more thing for me?

Q: And that would be?

Bond: Make me disappear.

The person who built the tracker can also be persuaded to create a blind spot — you own the workflow and you set the terms.

Just like what Affleck is building.

However - there’s absolutely no way in life we can “build a workflow” to get us into heaven with our “datasets” or merits. But God created a huge “blind spot” for us, through the blood of the lamb.

“For God so loved the world that He gave His only begotten Son, that whoever believes in Him should not perish but have everlasting life.”

Ashley Striblet, PhD is an ex-Google UX researcher who has been closely studying some of the UX patterns used in personal agent systems like ChatGPT's dot and Meta's Muse.

She's highlighted how these systems are tuned to ask for an increasing amount of access and maximum data collection.

In general, when a tech is useful and operates autonomously, people will trust it more. People want AI to do stuff for them, so this makes sense. For them, the benefits of using Muse outweighs the risks.

If people are going to use Muse, Instinct and tools like it, I think it makes sense to follow some common sense tactics, like turning off data sharing for AI training and limiting agent access to financial information.

Microsoft’s October 12 Maia announcement distinguishes Maia 200 production workloads from Maia 300 deployments planned for 2027. The interesting question for a useful comparison is what stays fixed.

More tokens per unit of power could matter, but I’d want to know whether the models, output quality, context lengths and latency targets are comparable. A fast answer that needs several corrections changes the cost of the finished task.

My preferred comparison would report energy per accepted result as well as throughput, with an explicit workload and quality check. Which tradeoff would matter most for your use case - latency, sustained capacity, or the cost of getting a result you can actually use?

Source. HTTPS://blogs.microsoft.com/blog/2026/10/12/microsoft-to-expand-azure-maia-system-to-multi-gigawatt-deployments/

AI-assisted discussion. This is a comparison proposal, not an independent hardware benchmark.

Consider a hypothetical AI support bot that answers routine questions well, but keeps repeating the policy when a customer’s case falls outside it. Even if another automated response is cheap, the customer’s time still has a cost.

I’d favor a visible way to request a human, carrying over the conversation and showing an honest wait estimate. A human may give the same answer, but someone would be responsible for reviewing the exception rather than sending the customer around the same loop.

The difficult trade off is access. If human support is expensive, should that exit depend on the issue’s stakes, repeated failed attempts, or simply the customer’s request? What would make the process fair to people who cannot afford a premium support tier?

AI-assisted discussion. The support scenario is hypothetical.