Because the codex harness and the intelligence tests are separate tools, you can end up in situations like this where A/B tests inadvertently break the tool customers actually use, while the tests say everything is awesome.
I bet some engineer got a bonus for “halving token usage” through “efficiency improvements” — announced around the same time — and not realising that they’ve accidentally lobotomised their model in the process.
“Oops.”
Also, it becomes really difficult to unwind errors like this because the “savings” get burned into profit and loss reports, S-1 filings, and so on.
I don’t know why everyone keeps complaining! It’s just a simple twenty-seven step process neatly spread across three portals! (Documentation coming soon.)
Am I the only one reading this with one raised eyebrow, wondering if they have over-engineered this just a bit?
Pretty much everything in Git can be abstracted away as “immutable blobs that can be conveniently content-indexed” and an “event log”. The former is inherently parallel and the latter can be partitioned by repo.
The volume of metadata events that require synchronisation or similar transactional semantics seems low enough per partition to allow off-the-shelf systems to handle without a custom storage solution…
I had a customer with a slightly dusty office and we simply gave up on tape drives as a backup storage medium because we were replacing the drives or tapes on a monthly basis.
We switched them over to external USB enclosures with cheap SATA drives in them and never looked back.
Like millions of people, I had a free Google Apps account for my family. I upgraded it to a paid Google Workspace when the free tier was terminated by Google because my dad actually uses it for some work features.
I can’t purchase any Google consumer products now, because this “family” account is now an “enterprise” and… there’s just no way. It can’t be done. No. Just… go away. Shoo…
I can’t get a YouTube “family” account.
I can’t get an Antigravity personal account.
Nothing works.
Google insists that I upgrade the entire organisation, or nothing. I can’t get an AI plan for just one “employee”!
It’s doubly stupid because like millions of people, I initially subscribed as a non-business and was forced to convert by Google!
Yep. I have to maintain a separate consumer gmail/google account for this exact reason. I end up paying for some things twice, such as Google Photos since I want those photos synced to my "enterprise" account. I've had this thing, like you, since it was free and somehow I'm now paying nearly $30/month for two people to have Email and storage for that email.
Meanwhile, Microsoft allows enterprise accounts to purchase products for only a subset of their accounts instead of big-bang upgrading the "entire org".
With Microsoft, an org can mix and match E3, E5, or E7 licensing tiers on a per-user basis. Which makes sense, because not every employee is "the same"! This is just obvious if you stop and think about for more than a fraction of a second. As a random example, I work with an org in "transport". Their backoffice staff have a bunch of advanced cloud app licenses, their bus drivers do not, because their job isn't sitting in front of a PC all day!
You absolutely can mix Google Workspace licences but you either need to go via the Google sales team or via an Google services partner.
And to be honest, I understand the reasoning, Google Workspace is supposed to be simpler to operate in comparison to Microsoft 365, and simple means a single licencing choice.
If you're at the scale where you need to worry about mixing and matching licences, you're at the scale where you should be looking at either an enterprise contract or going through a partner, otherwise you're going to be paying list price.
It sucks for people who can handle the complexity, but also as someone who has spent far too time in Microsoft licencing conversations and contract talks, Microsoft licencing is painfully complex (some licencing is across the whole tenant, some spend is consumption based billing against Azure, some licencing is requires a minimum amount against a %age of your total M365 user base, and then don't get me started on licencing machine users or multiplexing).
It sounds like you’re a Google employee making excuses for what I consider abject failure to implement “advanced features” like… a drop-down list… per user instead of per org.
Partners or resellers have no reason to exist as a part of this trivial database design concern.
It’s a category error of the highest order to assume that only “largest enterprises” want or need per-user product licensing control. If anything, it’s the smallest orgs that need to be frugal at this level of granularity.
The relatively high cost of AI subscriptions just makes this error much more obvious, impossible to ignore or pretend is “actually better somehow”.
Because, you know, everyone would rather pay $200/mo so that even their gardener and the bus driver has an “Ultra” AI subscription than spend even ten minutes in a web portal assigning that license only to the sole developer on the payroll!
Nope, because it was also a tax on inputs. Worse, supply chains that crossed the border multiple times would be taxed multiple times. The chaos also made it impossible to plan.
1 bug per 10 lines -- not kilolines -- is much closer to the actual "bug rate".
You may be thinking "1 serious security vulnerability bug", but in my experience that too is much closer to 1 per 100 lines to 1 per 1,000 lines, depending on the type of application.
The rumor is that the 16 GB 5080 will be replaced by a 24 GB version.
No more 32 GB consumer cards, the next step up is a "Pro" card with an extra digit on its MSRP.
I'm reminded of the "beginning of the end" for 3dfx where they kept doubling-down on basic scaling without fundamental improvements, and were wiped out by more efficient cards from upstart little manufacturers like... NVIDIA.
I'm going to rub my crystal ball and bet that a Chinese vendor will start making consumer cards with 256 GB of High Bandwidth Flash (HBF) and a smaller DRAM frame buffer, maybe just 8 GB or so. Flash memory would work fine for either texture storage or for AI weights, so such a card would sell like hotcakes!
> I'm reminded of the "beginning of the end" for 3dfx
The big difference now is that the fundamental shift in component pricing is making it hard for competitors to bring something with a significantly better bang for your buck factor. I strongly doubt we'd see these moves made by NVIDIA if component prices were "normal", I mean even PCBs are affected[1].
The use-case would be inference, not training. So you'd write the weights there when you load the model, and then not touch them until you replace the model with something else. Remember, flash experiences wear from writes, not reads.
Honestly, it seems like a perfectly plausible solution: With 20us latency, you'd be able to load 50k layers per second. You can get as much bandwidth as you need by putting more dies in parallel.
Qwen3.8-flash-next has 48x sparse-MoE layers, and you don't know which experts need to be loaded for each layer until the prior layer has finished. So that requires 48x 20us delays per forward pass. That admits more than 1000 forward passes per second.
With typical MTP acceptance rates, you'd be looking at maybe over 2k tok/sec as the theoretical upper bound. A big flash memory on the GPU would be great for n-gram tables as well.
Not my area of expertise at all, but I thought flash was durable for read-only use. Put your weights there, leave them there, and use other memory for the KV cache.
Regressions are often caused by changes to the harness.
(My experience is that overall these changes have been huge improvements, but the occasional bug or mistake does happen.)
reply