Cruxius

> they are absolutely using their newfound market power to enshittify their services at the moment. Citation needed, or are you one of the people who thinks that they're only pretending to be compute constrained?

Something Dario Amodei has been extremely consistent on is that advanced AI (potentially including recursive self-improvement) is going to act as a massive economic force multiplier, and that whoever gets there first is going to have a huge advantage. A lot of people take that to mean a first-mover advantage that’s basically insurmountable (i.e. 'winning'), even if he doesn’t usually state it that explicitly. That seems to be a big part of what’s behind his push to restrict China’s access to cutting-edge AI. From the outside it gets framed as a US vs China thing, but his own framing is much more about risk and alignment (i.e. that US-led AI ecosystems are more likely, in his view, to produce systems aligned with broadly good outcomes than alternatives). A lot of people claim he’s pushing for this for anticompetitive reasons, which feels like it ignores a lot of the safety/alignment work Anthropic has been publishing on the risks of poorly controlled systems. That said, until Elon Musk bestows his Neural Lace upon us we can’t actually read minds, so any claims about his “true” motivations are unfalsifiable either way. At that point it kind of just comes down to whether you take people at their word, and he’s been pretty consistent on this for years, 'when people tell you who they are, believe them' and all that. Sorry about that digression at the end, I'm pre-tilted by people who constantly claim Dario's just hates competition.

The first AGI might not be an LLM, but they're already a major part of the path to AGI, and a RSI LLM could get LLMs to the point where one could meaningfully assist with coding whatever AGI turns out to be.

The new one is to ask how many ‘r’s in ‘londonderry, to which they’ll confidently answer ‘3!’.

yes fwau?

I’m glad I got to see this post, and glad I got to see the high quality discussion around it that resulted from it being posted on this subreddit. If not here, then where could it be posted that would still give the same high quality discussion?

They claim that the hardware is open source too, so there’ll be a BoM and you’ll be able to order the parts yourself if you need to.

Any chance of Lora support?

Yeah, Microsoft tried making major changes under the hood with Vista and while they were necessary, the issues they caused permanently tarnished its reputation.

Yeah, I really miss the built-in temperature probe inputs on my old mobo which let me put a sensor either side of my radiator and control my fans by temperature delta. Having it built in is a great feature, since fan speed based on CPU temp is nonsensical for watercooling.

Based on how it bangs on about safety, my guess is cb-a is an Anthropic model.

I have an inverted case setup and it works similarly, CPU AIO pulling in from the front, fans on top blowing down onto the GPU intake fans.

whaaaaat? Regular human beings totally use the check emoji and number ~~their~~ our paragraphs.

Sounds like your Brain-1M model is running into some serious inference issues. The MoL (Mixture of Lobes) approach is novel, but based on your report, there are a few key bottlenecks: 1. Expert Lobe Activation Issues • The Frontal Expert Lobe (FEL) typically requires structured fine-tuning with real-world reinforcement learning (RWRL) rather than just pretraining on passive datasets. • You might need to improve its energy source (RTX 5090 was a pipe dream anyway—Frozen Food & Coke™ is a known unstable fuel mixture). • Consider a controlled sleep-wake cycle. The FEL tends to underperform when inference sessions extend beyond recommended uptime. 2. Hallucination Rate (33%) • Nighttime hallucinations suggest overactive default mode networks (DMN)—common in MoL models. • Mitigation strategies: • Increase physical activity (improves token coherence and reduces overfitting to irrelevant data). • Reduce caffeine-based clock-speed boosts, as these can cause misalignment in temporal processing units. • Optimize memory retrieval pathways through reflective journaling fine-tuning (a manual approach but effective in reducing drift). 3. MMLU Pro Performance Issues • Math-heavy tasks? MoL architectures often struggle with multi-step logic problems due to lazy computation allocation. • You might need to simulate retrieval-augmented reasoning (RAR) via external processing (e.g., consulting external knowledge bases or distributed compute nodes—aka “other humans”). • Consider implementing a low-latency meta-cognition layer (often built into MoL v2 via conscious reflection). 4. Hardware Constraints • While Frozen Food & Coke™ provide some baseline compute power, diverse nutrient intake could significantly improve processing speeds. • Memory expansion modules (Hydration & Sleep v2.0) can reduce random context drops. • If you can’t afford an RTX 5090, at least try to overclock with some regular exercise and daylight exposure. TL;DR: Fixing Brain-1M ✅ Activate the Frontal Expert Lobe with structured RL and real-world task repetition. ✅ Reduce hallucinations by managing energy intake and cycle resets. ✅ Improve MMLU Pro performance via external augmentation and structured recall. ✅ Upgrade hardware stability by balancing input sources (nutrition, rest, activity). Might not get you AGI, but at least you won’t blue-screen at midnight.

I think it’s standard across a lot of industries, I remember buying a pair of Bose headphones at launch and as I was paying for them the guy told me he was jealous since he wasn’t allowed to buy them himself for two weeks.

No one, it doesn’t ‘remember’ things from its training data. That entire part is a hallucination.

$24.10 AUD which is $15.15 USD according to the google, but most industries are also covered by what's called an 'Award Wage' which typically boost the minimum wage a bit higher. The price difference isn't quite as bad as it seems, I forgot that Aus prices include sales tax (10%), plus our consumer protections are excellent which adds another 5% or so to account for compliance costs, plus the extra cost to ship to a smaller market way in the middle of nowhere. They're still overcharging by a good $300 or so, but it's not the worst kick in the teeth.

Fuck me, a casual $800 Australia tax.

Even AMD did it, I still fondly remember my 4870 X2.

Brilliant, thanks for your help.

Okay perfect thanks, does that mean that if I don't care about PCIe 5.0 (it's unlikely that the 5090 will come close to saturating PCIe 4.0) I can grab basically any chipset and all I need to worry about is the specific mobo features I'm after (NIC, ports, RGB etc)?

The other piece of circumstantial evidence is that intel is refusing to confirm or deny there will be any future generations on this socket.

haha wouldn't it be funny if the bitlocker server where the keys are was also BSOD haha that would never happen

Introducing: BSOD As A Service (BSODaaS)! Tired of your computer running too smoothly? Want to relive the good ol' days of unexpected crashes and cryptic error messages? We've got just the thing for you! Welcome to BSODaaS: The Ultimate Disruption Solution. Why Choose BSODaaS? 🔵 Experience Authentic Nostalgia: Remember the thrill of a Blue Screen of Death (BSOD)? Now, you can have it on-demand, bringing back those nostalgic feelings of panic and frustration. 🔵 Unmatched Downtime: Ensure your productivity plummets to new depths. With BSODaaS, you'll get random crashes, ensuring you never complete a task on time. 🔵 Perfect for Procrastination: Need an excuse for missing deadlines? Look no further. BSODaaS provides the perfect alibi with its perfectly timed interruptions. 🔵 Customizable Errors: Choose from a variety of error messages. From "IRQL_NOT_LESS_OR_EQUAL" to "KERNEL_SECURITY_CHECK_FAILURE," we've got all the classics! 🔵 Compatibility: Works seamlessly across all platforms. Whether you're on Windows, Mac, or Linux, we guarantee a flawless fail. 🔵 Subscription Tiers: Basic Crash: Occasional BSODs, perfect for casual users. Pro Crash: Regular disruptions to keep you on your toes. Enterprise Crash: Non-stop chaos for entire office networks. Say goodbye to productivity! Testimonials: 👤 "Since we implemented BSODaaS, our team’s stress levels have skyrocketed. It’s like a time machine back to the 90s!" – A Happy Customer 👤 "I used to get my work done way too fast. Thanks to BSODaaS, I now have the perfect excuse to take long coffee breaks!" – Procrastination Enthusiast 👤 "Our IT department has never been busier. Thanks, BSODaaS!" – IT Manager Join the Revolution: Say goodbye to seamless performance and hello to delightful disruptions. Sign up for BSOD As A Service today and bring a little chaos back into your life. BSODaaS: Because who needs reliability anyway? 🚫 Disclaimer: BSODaaS is not responsible for any data loss, missed deadlines, or sudden career changes. Use at your own risk.

The difficulty of doing it aside, what’s the value? If it tells me it’s 95% sure it’s right how is that more or less useful than 50% or 80% or 99%? If accuracy matters then anything less than 100% is functionally useless, and if accuracy doesn’t matter then who cares how confident it is?

When Claude first launched on Poe it would often do that, but that made people mad so they ‘fixed’ it.

Might be where the LLMs are getting their incorrect answers from too.

no u

Are you seriously complaining about people picking an unrepresentative photo and following it up with professional shots taken with studio lighting, professional makeup and edited in photoshop? It’s very possible she looks fine in a photo which isn’t chosen to portray her in either the worst or best possible light, but that’s not something you’ve provided any evidence for.

It typically does, but not for very large values of 2.

How's that net neutrality going for you?

I dub thee the godfather of whinging on reddit.