vient

Author mentions trying incognito mode without success.

I also tried their strictly mathematical problem description and got filtered 5/5 times.

Interesting, I thought that it can't be right, Fable can't refuse to answer a strictly mathematical problem — well, 5/5 attempts did switch to Opus. Amusingly, one attempt spent almost 10 minutes thinking how to prove NP-hardness only to abruptly switch.

Wonder if AMD MI350P release will affect setups like this. From what I've heard, the price will be pretty similar to RTX PRO 6000 while having 50% more VRAM which is additionally an HBM3E instead of GDDR7.

At least this time they actually created an incident https://slack-status.com/2026-05/46d9d5d41fbedd9f

https://slack-status.com does not report anything so far. We observe issues sending messages via app and API.

Edit: incident appeared https://slack-status.com/2026-05/46d9d5d41fbedd9f

A lot of 'premium' branded instant coffee is ~£42/kg. That's £3/kg more than my premium, locally-roasted, single-estate Colombian coffee beans.

You need 7-10 times less instant coffee to make a cup though, so your beans are a lot more expensive in the end. From my experience, cup of coffee is either 2 grams of Nescafe Gold or 16 grams of beans.

I assumed that AI features require server, so the server knows how many users it has. If some features are local then it's not so straightforward.

Approximately all people update Firefox so you don't need telemetry to count "AI disabled" installations, instead you can derive it as "updates requested" minus "AI enabled".

On Windows it is Alt+0151. Harder to use than on Mac but definitely possible, I frequently use it.

On recent versions Shift+Win+- also work, and Win+- produces en dash.

Peculiar that in IMPLEMENTATION_NOTES.md Claude thinks it is 2024 and not 2026 (see Work Log)

I wonder if Windows and Linux just can't yet work on heterogeneous CPUs as well as macOS does. Intel chose an interesting direction here, going straight from one to three kinds of cores in one chip. I almost never see LPE cores being used on Windows, and on Linux you have obscure soft like Intel LPMD which I tried but was not able to notice any battery life improvements.

You can easily download unsigned binaries and run them

Of course, but I assume you don't really need to install third-party apps to control hardware. In my case Alienware and Dell bloat came from me setting up an Alienware monitor. MSI bloat came from setting up MSI GPU. Intel Killer stuff just got automatically installed by Windows Update, it seems.

Microsoft Defender

This one I immediately disable after Windows installation so no problems here :)

On work we get CrowdStrike Falcon, it seems pretty tame for now. Guess it depends on IT-controlled scan settings though.

building a CPU that doesn't guzzle battery

It may be the software problem as well. On Windows I regularly need to find which new app started to eat battery like crazy. Usually it ends up being something third-party related to hardware, like Alienware app constantly making WMI requests (high CPU usage of svchost.exe hosting a WMI provider, disabling Alienware service helped), Intel Killer Wi-Fi software doing something when I did not even know it was installed on my PC (disabling all related services helped), Dell apps doing something, MSI apps doing something... you get the idea.

It seems like a class of problems which you simply can't have on macOS because of closed ecosystem.

Without all this stuff my Intel 155H works pretty decently, although I'm sure it is far away from M-series in terms of performance.

The problem here is not with Optimus though.

git-archive downloads only strictly necessary files but is not universally supported

https://git-scm.com/docs/git-archive

Same can be done using git and tar

    mkdir -p <out_dir> && git archive --remote=<remote> --format=tar.gz <branch> <files...> | tar -xzC <out_dir>

For reference, I get 29 tokens/s with the same model using 12 threads on AMD 9950X3D. Guess it is 2x faster because AVX-512 is 2x faster on Zen 5, roughly speaking. Somewhat unexpectedly, increasing number of threads decreases performance, 16 threads already perform slightly worse and with 32 threads I only get 26.5 tokens/s.

On 5090 same model produces ~170 tokens/s.

Won't say I got much longer battery life, and even what I got may be as well explained as "TLP made energy profile management almost as good as on Windows, and then Windows's tendency to get a bunch of junk processes seeping on your battery tipped the scales to favor Linux". Also I ended up switching back to Windows because of never-ending hardware issues with Linux, installing it on 155H back in February 2024 was especially rough but even 6 months later I randomly got Bluetooth not working anymore after Ubuntu update.

My TLP and LPMD configs: https://gist.github.com/vient/f8448d56c1191bf6280122e7389fc1...

TLP: don't remember details now, as I recall scaling governor does not do anything on modern CPUs when energy perf policy is used. CPU_MAX_PERF_ON_BAT=30 seems to be crucial for battery savings, sacrificing performance (not too much for everyday use really) for joules in battery. CPU_HWP_DYN_BOOST_ON_BAT=0 further prohibits using turbo on battery, just in case.

LPMD: again, did not use it much in the end so not sure what even is written in this config. May need additional care to run alongside TLP.

Also, I used these boot parameters. For performance, I think, beneficial one are *mitigations, nohz_full, rcu*

    quiet splash sysrq_always_enabled=1 mitigations=off i915.mitigations=off transparent_hugepage=always iommu=pt intel_iommu=on nohz_full=all rcu_nocbs=all rcutree.enable_rcu_lazy=1 rcupdate.rcu_expedited=1 cryptomgr.notests no_timer_check noreplace-smp page_alloc.shuffle=1 tsc=reliable

TFLOPs are not the same between architectures.

Shouldn't they be the same if we are speaking about same precision? For example, [0] shows M4 Max 17 TFLOPS FP32 vs MAX+ 395 29.7 TPLOFS FP32 - not sure what exact operation was measured but at least it should be the same operation. Hard to make definitive statements without access to both machines.

[0] https://www.cpu-monkey.com/en/compare_cpu-apple_m4_max_16_cp...

Something is wrong with power governor then. I have an opposite experience, was able to tune Linux on a Core Ultra 155H laptop so it works longer than Windows one. Needed to use kernel 6.11+ and TLP [0] with pretty aggressive energy saving settings. Also played a bit with Intel LPMD [1] but did not notice much improvement.

[0] https://github.com/linrunner/TLP

[1] https://github.com/intel/intel-lpmd

Weren't we comparing CPUs though? Those Blender benchmarks are for GPUs.

Here is M4 Max CPU https://opendata.blender.org/devices/Apple%20M4%20Max/ - median score 475

Ryzen MAX+ PRO 395 shows median score 448 (can't link because the site does not seem to cope well with + or / in product names)

Resulting in M4 winning by 6%

Oh, I see - hard refresh consistently shows HTTP/2 but after one or two soft refreshes it becomes HTTP/3 for me until next hard refresh.

Edit: it is always second soft refresh for me that starts showing HTTP/3. Computers work in mysterious ways sometimes.

Limited to macOS? Does not reproduce in FF 141 and 142 on Windows.

AMX is indeed a very strong feature for AI. I've compared Ryzen 9950X with w7-2495X using single-thread inference of some fp32/bf16 neural networks, and while Zen 5 is clearly better than Zen 4, Xeon is still a lot faster even considering that its frequency is almost 1GHz less.

Now, if we say "Zen5 is the leading consumer CPU for AI" then no objections can be made, consumer Intel models do not even support AVX-512.

Also, note that for inference they compare with Xeon 8592+ which is the top Emerald Rapids model. Not sure if comparison with Granite Rapids would have been more appropriate but they surely dodged the AMX bullet by testing FP32 precision instead of BF16.

True, but that has nothing to do with tagged pointers.

Memory page size should be transparent for tagged pointers (any pointers, really), I don't see how they can be affected. You have an object at address 0xAB0BA, does the size of underlying page matter?

Would also be nice to remove empty categories from tree view. For example, right now you can uncheck VSX and still see "Memory Operations - VSX Unaligned ..." full of empty tags.

But what's the point in acknowledging numerical issues outside of [-1,1] if polynomials do not even work there, as author explicitly notes?

Kalimba, VE

No idea what this is, and Google won’t help me.

Seems that Kalimba is a DSP, originally by CSR and now by Qualcomm. CSR8640 is using it, for example https://www.qualcomm.com/products/internet-of-things/consume...

VE is harder to find with such short name.

They have some NVIDIA support in the form of external project: https://github.com/openvinotoolkit/openvino_contrib/tree/mas...