Those resets and the removal of the 5-hour usage limit are quietly anchoring me to a much higher usage baseline. I've stopped rationing and just spawn a bunch of agents to work at whatever pace I want, because there always seem to be more resets on the way (at least true for the last week). And now I am actually worried about that if one day they just stop doing so, my "normal" workflow will suddenly exceed the limit, and upgrading will feel like a step backwards.
Can anyone explain how you “win” the market of super intelligence? Particularly with open weights models now rivaling the frontier, it seems like a race to the bottom even if the prices don’t yet reflect that.
"race to the bottom" is the negative framing of "competitive market prices".
So a company or union might say "this is a race to the bottom" when someone new enters their market, but to people buying their services this might be seen as welcome competition.
Do you actually see a negative impact from competition in this area? Or do you just mean competition will further reduce prices?
"race to the bottom" is also a response to a "market for lemons". It's not necessarily a good thing because while pricing drops to the floor so also does value to the customer in general. Usually it happens when price is very visible but the details of what the buyer actually receives are not.
Regulatory capture. You get them to outlaw the part of the competion (safety!) that is unwilling to pricefix and participate in your margin and market division agreements.
Don’t think of this as super intelligence. Think of it as vendor lock-in.
We have a good compare, which is cloud in the early 2000s.
Everyone (at the time) thought cloud compute costs would go to zero.
What most corporations didn’t realize is how entrenched your workflows and processes get when you adopt cloud and you become heavily locked in that ecosystem.
That same ecosystem lock-in is what the frontier labs are hoping for with AI.
You patent and protect (as best you can) the missing ingredients needed to get to AGI. Only half joking and also scary to contemplate. Qualcomm’s CDMA patent is on example. ARM and Texas Instruments two other examples.
Everyone is raising the bottom. Kimi got 60% more expensive during the 2.x cycle despite staying the exact same size.
Now K3 is almost 6x the cost of the original K2 checkpoint, and while the parameter count finally jumped, it's still an extremely sparse MoE and definitely does not cost 6x what the original K2 checkpoint did to host at scale.
Race to the bottom only takes real effect when there's a cap to the capabilities, otherwise everyone races to the bottom of a rising target (how economically valuable the tokens are)
"Why" as in, why take lower margins when Moonshot currently can't service all the demand for the model anyways. Based on past models no one is going to massively undercut Moonshot: few have the chops to serve it as efficiently as Moonshot and of those few, most of them don't go for being the cheapest, they go for being fast + reliable (think Together, Fireworks).
You get what you pay for applies very much with how many axes there are to serving these increasingly large models.
-
And for "with what compute": as the value of a token goes up, what people are willing to pay for compute is going up.
Every once in a while I'll see a story about falling rental rates, but with even slightly more established clouds I've been seeing availability get worse and worse over time.
I'm pretty sure the only reason the highly informal indexes don't reflect this is because every neocloud trying to cash in on an NVIDIA Inception discount kicks off by selling unrealistically cheap compute for a bit.
AGI will be insanely priced. LLMs are retarded childrin compared to proper intelligence. So this is just the entry level intelligence like thing. But I doubt there will be an AGI accessible by anyone.
If AGI comes to exist it won't be "priced" at all, since the lab that creates it will either quickly be seized by the gov't for national security, or they will become the most powerful organization in the world and have no need to sell services to other corporations. they will become the only corporation.
Good sci fi premise, but not at all how AGI will happen.
It’s not going to be a singularity at one moment of time. It’s not going to be instant runaway self-improvement, no matter what doomers and fetishists say.
It’s going to be gradual. We’ll see glimmers of AGI, and the “G” part will be about gradual broadening of domains and deepening of capabilities.
All of the coding harnesses are already using their own tools to self-improve, and the HITL component is getting less frequent and at higher levels of abstraction.
That’s how AGI gets here: very gradually, no hard takeoff, and nobody will be able to pinpoint when exactly it happened.
So: also no single lab with a massive advantage, no government takeovers. It’ll be a lot less dramatic than the extremes believe. IMO, of course.
The problem is... the moment someone gets it and it really solves hard problems and solves scifi level shenanigans, the given country that owns it, could gain such unfathomable lead above anyone else that it will almost surely lead to an all out war. I don't even know whether it is possible to conceal that you have such capability....
Imagine all the fear- and warmongering kingmakers and powerful individuals when they realize they have no power over anything or anyone.....so game over for them. They won't like it at all, at all.
Another problem, that in order to make it understand real life, it needs robots or humans wired into it (brain interfaces) in order to test certain things in the real world. And that is another level we know almost nothing about, at least on the surface.
Someone could have a small private breakthrough tomorrow that gives sample learning efficiency of the brain, online learning, and consolidated memories.
If it’s 50 years and gradual, there will be many places very close. The only “winner takes all” scenario is a sudden breakthrough when nobody else is close.
If gradually one of these labs ends up with the top model that makes all the others uncompetitive, what's the difference? Why do you see time as important if the end result is the same? Winner still takes all, no?
This is actually what I've started to feel a vague nagging concern about as well..
Say what you will, there's no way I'm going back to non-AI assisted coding. Even though I don't use AI to generate code or assets, it's great for reviews and brainstorming etc.
What if OpenAI/Anthropic decide to do a Netflix/Spotify move and pull the rug out from under us one day?
Like skyrocketing the price, or limiting peasants to older models (because Glorious Leader said so), or maybe some leak comes out that they've been spying on us all along.
For context, Phi is an AI-powered Chromium browser for macOS. I think many people here may not be familiar with it. The closest reference point is probably Arc, although Phi's goal is slightly different: a browser that you and your agents can both use.
The main changes in version 2.0 are Spaces for organizing tabs and fully sandboxed Profiles, which keep cookies and sessions separate. In practice, that means things like multiple Google logins no longer collide. URL rules can also ensure that specific sites open in the correct space/profile.
The assistant now has whole-browser context, and the guardrail for removing sensitive data runs locally on your Mac instead of on a server. I recently worked on optimizing the local model path on the Apple Neural Engine because it seemed wasteful to leave that hardware unused.
I am also experimenting with a new memory visualization system (we call it Nebula). Rather than displaying browser memory as a list or force graph, it presents it as a spatial surface that you can explore over time. This feature is still evolving, but it has been interesting to work on.
Next, we want to enable agents such as Claude Code or Codex to control your actual browser window rather than a separate headless Chrome instance.
Got myself the $20 subscription and tried it out. The 5-hour limit runs out surprisingly fast. Quality is okay but it feels slow, and even with my $20 Claude subscription on Fable, the credit usage ends up being lower. Fable usually catches issues in my Opus 4.8-generated code that I'd miss otherwise, but Fugu didn't. Makes me wonder if it's really at the Fable level. Hard to see the value here.
I use my Flipper Zero weekly (or more frequent). This new model feels much more powerful than the ones based on RPi Zero as a handheld device. I like how they managed to include two RJ45 ports and a USB-A port for connectivity. However, it's still too bulky for me. Perhaps when I get one, I'll try carrying it around all day to see how it goes.
There's also a nano SIM slot. With the two Ethernet ports, it's perfect for use as a mobile router. This use case alone is good enough for me.
For such a powerful device, I think the lack of a QWERTY keyboard and the inherited orange backlit monochrome display are two of its shortcomings. I don't want to carry a keyboard or screen with me, I want it to be able to take more human input/output without accessories.
For those interested in hackable, handheld Linux devices, the M5Stack Cardputer Zero is also worth a look. It will launch on Kickstarter soon, and I have reserved an early bird spot.
I'm curious what this means for ChromiumOS and downstreams like FydeOS.
If Google is now pushing this "intelligence‑first" desktop experience, how much of that work is likely to stay in the proprietary ChromeOS/Googlebook layer vs. land in upstream ChromiumOS?
The OS on these Googlebooks will probably be a lot closer to Android 17 than to current ChromiumOS. Google has been consistent in saying that they're phasing out the ChromiumOS code base (while continuing the support the Chromebooks they've already sold) in favor of modifying AOSP to work better on laptops and desktops.
I went a slightly different route. My switches are linked with 10Gbps SFP+ across the apartment, but it was way too late (and too much hassle) to pull proper in-wall fiber. Instead I used one of those ultra-thin (around 0.1mm), unshielded fiber cables, and just snaked it through door frames and taped it along the walls. I'm genuinely impressed this stuff exists, it makes retrofitting fiber into a finished space so much less painful.
Most of my edge devices are still on 2.5GbE though, and I'm increasingly aware that for anything with plain SATA disks, the drives are the real bottleneck. Once I LAG'd 2×2.5GbE to get a 5Gbps pipe, it became obvious the network wasn't the slow part anymore in a lot of cases.
And yeah, the 10GbE SFP+ modules run hot, so hot that I would not lay my fingers on them for more than 2 seconds. I stuck 2 copper heatsinks on my module, not sure they do much but the module runs smoothly. Even so, I'm pretty happy with the overall setup: from my 10GbE-equipped Mac I can saturate multiple machines at once and I no longer think about the network most of the time, which was the goal.
And after getting 10Gbps working at home I was getting greedy, and looked at InfiniBand as well, 40Gbps and proper RDMA is very tempting compared to Ethernet. The catch for me was the practical side: IB needs PCIe slots and those chunky, inflexible cables. With most of my stuff being laptops, mini PCs and Macs, I just couldn’t see a clean way to route those or even plug cards in everywhere, so in the end the "door-frame‑friendly" skinny SFP+ fiber still won out for this apartment.
Good catch — that was actually my mistake. I mixed up cm and mm from the marketing material, and when I actually measured just now it came out to around 0.9mm. So not quite as thin as I claimed, and I probably wouldn't have noticed if you hadn't asked.
I did find some Kevlar-reinforced options that are supposedly ~0.3mm, but they seem to be raw fiber without connectors, purposed for drones, and I'm not sure about global availability.
Ah, I see. Thanks for checking. I'm considering doing a similar run at home, so was curious if there were any options that thin. 0.9mm should be plenty to work with. :)
Yes, they are with one GPU core fused off. I came across a die shot on the internet[0], and the GPU cores look huge. With some rough calculation, I estimated that the GPU cores together take up about 15.5% of the die area.
I don't know much about photolithography, but I assume the same percentage of single-defect dies could be limited to a single GPU core failure, it's actually pretty surprising that Apple can get enough of these "rare draw" chips to build and ship a real product.
If the shortage continues, I would expect that they start using fully functional A18 Pro chips with one GPU core disabled with software. It kind of reminds me of the AMD Athlon days when user could use a pencil to unlock extra cores.
I don't see my first GPU on there, it was the humble GeForce4 MX440. It could run almost any game I cared about for a surprisingly long time, even if it's not a true modern card.
These days almost all my machines are on iGPUs baked into the CPU. There's way less fun for me, but they are a lot more compact at least.
The GeForce 4 generation as a whole, while being solid enough cards, were historically not interesting. They were just basic spec bumps over the GeForce 3. No new features or similar. And, critically, the 9700 Pro released the same year as the GeForce 4 and absolutely smoked the living shit out of it.
The MX440 allowed players that were playing games on id Tech 3 to finally play at high frame rates. I remember this card being all the rage back then in pro gaming circles for this reason.
The MX440 was an entry level budget card? If it was all the rage in pro gaming circles at the time that's really just a reflection of how poor pro gamers were back then rather than anything to do with the MX440 being particularly noteworthy. In fact looking back at old reviews, it was if anything a flop. Launch MSRP was too expensive for the performance it offered. Especially when it was a DX7 card surrounded by DX8 cards at almost the same price point (including Nvidia's own Ti4200 for just $50 more)
I'm on a 3060 currently and the changes in the 4xxx and 5xxx just aren't appealing to me. As soon as iGPUs get 3060 performance I'll probably switch. And they aren't far off.
I'm puzzled by Espressif's naming here. We had the ESP32-S3, so "S31" sounds like "S3, variant 1," but this part doesn't really look like a simple S3 variant. And then there's an ESP32-E22, but no E21 or even a plain E2 anywhere.
It was because IA-64 was a completely different unrelated architecture that until AMD succeeded with K8 was "the plan" for both 64bit intel roadmap and the roadmap to kill off compatible vendors (AMD, VIA)
Got my RPi 5 16GB quite a while ago for around $160 and already thought that was expensive... It’s still powerful enough for almost everything I throw at it, honestly a bit overkill in most scenarios.
With prices steadily going up, for me it's starting to feel more sensible to repurpose the RAM sticks I've collected from old PC builds / laptops and just throw together small amd64 boxes instead of buying more RPis.
I wonder if there are low power Intel or AMD boards that accept DDR3. So many sticks of 2 / 4 / 8GB DDR3 inside laptops going into recycling or landfills which would do perfectly fine for low power purposes. Hell, performance for standard workloads scales with access times, not bandwidth, and DDR3 sits nicely at CAS8 1600MHz and CAS10 2133MHz..
