Instead, Export Administration Regulations (EAR) applies to exports of technology (software and hardware). Read more here: https://en.wikipedia.org/wiki/Export_Administration_Regulati...
About MegaSpeed: They were NVidia largest customer in Southeast Asia (which includes Singapore), buying about 140,000 GPUs in the last 2-3 years. Then, then built data centers in Southeast Asia and rented the GPUs to Mainland Chinese companies. To be clear: From 2023-2025 (their peak), this business model was not prohibited by United States GPU export controls. However, that will probably change soon.
This isn't to discount the issues with NVIDIA and whether they knowingly participated in circumventing chip bans. Or conflict of interest issues. But running a business that benefits China as a Chinese national and making some reports to the nation just doesn't seem like "spy" to me any more than being part of the British East India company made a Brit a spy.
But Jensen Huang has admitted to a longstanding affair with the CEO of a company that appears to exist only to evade ITAR restrictions. It's just impossible to believe that Jensen did not know what MegaSpeed was doing. While by all appearances he facilitated and benefited from it personally.
And one can also believe that ITAR restrictions on GPU sales are dumb, but evading them still counts as selling sensitive weapons to the enemy, both legally and politically.
My understanding is, the thing that got him effectively fired, [0] is not the affairs themselves, it was not being candid with his board about the extent of the problem–he admitted to them the affair with Alice Huang, but didn't tell them about his other affairs (including at least one with a direct report) until after the board was made aware of them independently.
A board can put up with a CEO who cheats on his wife. It can't put up with a CEO who hides things from it.
[0] formally speaking he resigned, but it seems likely the board communicated to him "if you don't resign, we'll fire you"
https://www.cnn.com/2026/11/12/politics/alice-huang-us-intel...
Some were allegedly with direct reports. One was allegeldy with an individual who posed a security risk. In each instance, there significant risk to the board, the company and shareholders.
He's obviously been very succesful in many elements of his role but this is very poor decision making. He could have just had affairs outside of work and been fine.
I think if it was one affair, this would hold. It is allegedly 3 and counting, and includes a likely spy from a rival nation, and likely knowledge of selling to the rival nation despite the current laws.
You can think he is amazing at many parts of this job, but at this point it is less of a mistake, and more a pattern of behavior that he knew was hugely risky and decided to cover up.
This is a terrible time for Nvidia employees. But the solution is not to invite the cause of the terrible time back in.
They've also taken a pretty big pay cut this week and potentially into the future.
At my day job, we have a calendar integration that polls Microsoft's Graph API across hundreds of Microsoft 365 tenants. We saw our first metric spike in 5xx errors at 2026/11/12 17:33 UTC. As of writing it's 2026/11/13 09:44 UTC and we are still encountering a full outage, so 16 hours and 11 minutes.
With a SLA calculator [1], assuming no other downtime this month and entering 16h11m, we get:
* Yearly SLA: 99.82%
* Monthly SLA: 97.78%
Next, what is the actual agreement? I found an SLA document list here [2].
I downloaded Service Level Agreement for Microsoft Online Services (WW) English (2026/10/01). There's a heading for Exchange Online that says:
Uptime Percentage : Service Credit
< 99.9% : 25%
< 99% : 50%
< 95% : 100%
Which seems to agree with what the original commenter said for this month.
[2] https://www.microsoft.com/licensing/docs/view/Service-Level-...
I started digging into the historical blog posts, GitHub comments, etc. and I realised everybody at MS who'd been working on them, SWEs, PMs, architects, etc. had either left the company, or moved to other teams. And the one constant about the internal moves was that people were now working in some "AI" related project or product. Literally, I didn't find a single person who wasn't now working on AI.
Clearly, there has been a huge internal push to get into AI. Resources have been reallocated, and anything non-AI related has been starved. And since working on an AI project is probably considered the only route to advancing your career, anybody with any ambition has moved to a team doing something AI-related. The result is that huge numbers of products and teams doing vital but non-AI related work have been hollowed out, and now we're seeing the consequences.
I don't expect a fundamental shift but the boss having no access to his mails certainly seems like something I would bring up quite a bit.
BBC is running live coverage of it at the level of government and ministers, feels surreal
Microsoft has a moat, but their moat is having billions of dollars to spend on data centers. It's a failure everywhere else. It'll continue to rake in billions in revenue because people need compute now, but it's rotten at the top.
WinUI 3.0, WinAppSDK, and related CsWinRT, C++/WinRT are in a ridiculous position six years after Project Reunion announcement, due to ever meagre resources.
XBox app keeps under delivering user experience, and bugs on patch Tuesdays are getting a joke by now.
All this in a four trillion valued company.
It's just a wild theory, but maybe there are still too much employees left, standing in the way of AI? I would try firing another 60,000 first.
What could go wrong, said all of the hyperscalers. Enjoy your multi day downtimes.
6 hours ago they seemed to not exactly understand what the problem was, and that was 6 hours into the outage.
This is concerning.
I get that some specialisation is useful, but I would like to see more T-shaped data folks.
https://news.ycombinator.com/item?id=49833842
But to explain further, Gnome is run by people who have the attention span of a dead goldfish. They're constantly changing things, a good example is the control centre where it used to be just like other systems but in Gnome 3 they made it into a wall of incoherent text and disorganised categories.
They often tend to ditch things halfway and create new ones, Evince -> Papers, Totem -> showtime, EOG to Loupe, Cheese to snapshot, and so on and so on.
And for decades, they kept removing much used features and configurations and broke things and everytime you search for a fix, there was 10 version and a new one each week because Gnome keeps changing things.
In short, Gnome is what happens when you think you're Apple but have the design knowhow and depth of a paddle you couldn't drown a fruit fly in.
The hot patching is unusable. It takes 11s to compile the patch after a simple change: (Some("hi") -> None), and then relax gets a compiler error when refreshing the page.
The project where it miscompiles only has a few very basic dependencies; a clean compile takes 1.3s, incremental just 0.25s and tests take ~0.1s. Relax takes a whole 1.4s to run the tests (unsuccessfully).
If GPUs are going to be priced out of consumer reach anyway, I would rather see the GPUs used to create open weight models than just to make these AI companies richer.
“In facilities across the region, Megaspeed rents Nvidia chips to Chinese tech giants such as Alibaba, an arrangement that does not violate current US export control regulations.” See [1].
[1] https://www.atlanticcouncil.org/dispatches/beyond-the-spy-sc...
I have no idea whether this is legit or not (there's no link to any code), I'm mentioning this because all the comments here are assuming truth without verifying the contents.
Apple has anti-spam app policies that basically limit innovation and competition in certain product categories. An extreme example would be Apple not allowing any more "sleep music apps" b/c there is enough in the store.
An investor actually approached me about buying a couple mobile apps I made b/c they are "aged" and don't need to go through the rigorous first-review approval process
But a new Flash model is very welcome. Argon and Carbon are very slow. In the time it takes for them to complete a single task, Opus already managed to complete multiple things. They’re so slow, I started using 3.8 Flash over them. (A lot of it is also because of how terrible agy is. Before doing any task, Gemini feels the need to hover up a bazillion files into its context when the task is something simple like “update the docs of this controller”)
Also this sounds like Argon is dead now
It hard to say they are doing this now because they need the money now - they are committed to spending close to $1T ($1000B) on datacenters/etc, and the IPO will only raise them around $50B (5% of what they need).
The timing seems more decided by the fact that growth may never look better than it does now as tokenmaxxing gives way to customers looking for ROI and cheaper alternatives to frontier LLMs, such as decision models and non-frontier LLMs which are rapidly becoming good enough for many use cases.
- rsi won't happen immediately, and is inherently unpredictable, so they need to bridge the gap
- it makes it easier to trade their shares/ownership with their suppliers. they're fighting over compute allocation like everyone else.
- their "public benefit" plans have always involved trying to share the upside with the public before it goes "exponential"
They looted all human knowledge, will displace jobs, they're disposable themselves, and no strong moat.
Why is such a company even allowed to be IPOd?
Because private rounds still dilute you.
And VCs tap out well before $100B.
And they'll need more than $100B to accelerate widespread economic diffusion.
And RSI can't hallucinate gigawatt power plants and TSMC fabs into existence.
Those 3 days are typically followed by an almost infinite number of bug cycles until it gets so bad you basically redo it.
Are they? I'm on a macbook M5 pro 48GB (so, much better hardware than most users), and the speed alone is a problem, qwen 3.8 27b is ~10x slower than opus 5.5.
Capability-wise... For software engineering, qwen is nowhere near as good as opus 5.5. Might be fine if your usage of LLMs is close oversight, but it's not how I use them (I do review AI work closely and iterate, but I expect agents to autonomously do good work nowadays). I've not tested it much for "life assistant"-type things, but the speed would make it unbearable.
For some reason they need the $100B today at the cost of dilution, rather than additional private rounds and/or using cashflow+profits to continue the advancement.
To me, this technology and these AI companies, in their current form, are as far away from the 19th century landscape paintings as they could get. As if they're trying to leech off them and fix their branding/message. Doesn't seem fair for some reason
Edit: Painting identification, my bad
I don't think this is entirely true. Companies are deploying open-source LLM gateways to be able to do that very easily now. Today, you can switch Claude CLI to any model besides Anthropic via https://github.com/BerriAI/litellm and there are major companies doing that.
This works with your favourite ADEs such as Orca as well so I disagree with this premise. The whole experience of switching is pretty seemless based on personal experience.
For society more broadly, what I’ve seen is a bunch of white collar workers under threat (real or perceived) of losing their livelihood, and a huge uptick in the amount of interference I’m having to run as my family members get taken in by a wave of increasingly sophisticated scams and phishing attacks.
Edit: I should clarify that I don’t think this justifies their valuation. It puts them in competition with much more entrenched players with established customer bases, and who aren’t in massive debt to provide their compute. I’d also like to distinguish between a model provider as someone who provides weights vs a model provider who connects their GPUs and models to your codebase and internal systems - I think the latter is where the sustainable business is.
Incidentally this is why I still feel Google is well-positioned in the long-term AI race. If they can remove their head from their ahh and make a coherent product offering across Workspace, GCP and Gemini, they could become as entrenched as MS.
I would assume AWS is probably the most interested in acquiring Anthropic out of the three major cloud providers.
1. Price range: $70–78 with midpoint at around $74.00
2. Gross deal size: $85.3B (IPO + $26B private placement)
3. Implied market cap: ~$968B (at $74, 13.09B shares)
4. Q3'26 revenue: $17.0B (+1,127% YoY)
5. Run-rate revenue: ~$81B (as of Oct 2026)
6. Compute commitments: $900.9B (through 2045)
7. Ticker: ANT (Nasdaq)
8. Customers with ≥$100K, ≥$1M and ≥$10M TTM revenue: 11,900+, 1,900+ and ~250.
9. Customers under $100K: ~40 million
10. Fortune 500 penetration: over 50%
11. Global 2000 penetration: over 30%
Anthropic and OpenAI will have to become more of the end user application stack. Files, calendars, emails, photos, and much more have to pass through them for them to keep making money. Else they cannot make money like the way they have been in this early era of LLMs.
What I want to know is whether someone has basic DSA understanding. Is the person familiar with any programming language. Can they translate business requirements into code? Can they disambiguate weird things PMs say into a viable plan?
You do not get a better signal about that from asking harder leetcode problems.
Never once in my career has writing a tricky algorithm under time pressure been relevant. The closest thing I can think of is debugging complex systems in an outage, but then it’s more like familiarity with bash commands and general debug skills.
As for LLMs and coding, companies need to figure out how to differentiate good and bad outputs here and what lasting engineer inputs to select for. And I don’t have any idea about this.
There are a lot of people telling you how they interview and why. Almost none of them have shown even the slightly indication that they are aware there there is scientific research on what works for an interview, much less that they apply it. They also are not showing any indication that they have done any testing - to see if their questions help hire better people.
If I said they could get equal results by picking resumes at random and hiring them without any other screening could they refute that? If so have they justified every part of their process or just one step that is useful and the rest they don't know?
we also ask the candidate to open a PR against the repo just to see if they know the basics of git
That's the basics of Github, not git.A candidate may be able to open a PR and not be able to make a commit without an IDE.
Another candidate might know plenty of git, but haven't worked with Github.
In general - I am wondering how engineering interviews change in the age of LLMs, Jane Street, tech-wise is a "thought leader", at least having opinions, sharing them is part of how they recruit "the best", i think oxide does this too.
So in some way, in this time of change it makes sense that they are noisy... It also makes sense that the noises they make may filter down...
Today: https://news.ycombinator.com/item?id=50467297
Yesterday: https://news.ycombinator.com/item?id=50455394
Two days Ago: https://news.ycombinator.com/item?id=50438261
It's a bit odd (?!).
Probably just random fluctuation.
first, a coding problem that is solved live (right now, online) -- it's not a hard one and most LLMs can easily one shot it.
BUT: if you are not a senior developer, you need to do it without LLM, and if you are a senior+ and do it via LLM, the bar is super raised and now you need to explain in details everything the LLM wrote as if you wrote it yourself.
the main component here is to know if the candidate understands coding in general, understands problem solving and how to dive into an existing (although quite small) codebase. solving the problem is not really the goal -- in fact, most candidates we hired did not solve it 100%, but were able to understand and explain everything that they put it on code. this usually takes 1h30.
(we also ask the candidate to open a PR against the repo just to see if they know the basics of git -- you would be surprise how many people cannot do it).
the second is an architecture interview, where we build a system together BUT instead of doing the whole spec, miro board, whatever, we do it in more of a free way: we start with a really simple, MVP level, spec and we ask how the candidate would build it. then, we add another feature, and we ask how they would go for it. during this whole thing, i ask random questions like "why would you use x instead of y?" or "why do you prefer this approach?".
one important thing: we allow the candidate to say "i don't know" and we tell the candidate that there's no gotcha questions (which is true, i hate these kind of questions). this lets the candidate be super comfortable (feedback we had from real people).
in general, this flow helped us a ton and i really really like running it :)
(edit: spelling)
This thing was actually found despite its location, not because of it. If it had been in the opposite part of the sky we almost certainly would have spotted it sooner, given that it's one of the brightest TNOs we've discovered.
We're finding it now because Vera Rubin (the instrument that found it) just came online and is exceptionally good at finding these kinds of objects specifically, even when they're hiding amidst a dense population of stationary background stars.
[1] https://edition.cnn.com/2026/10/28/science/ed138-dwarf-plane...
Classifications impose structure on a collection of celestial objects, and we can derive relationships, commonalities, understand their development, the physics governing these objects, and build useful physical and mathematical models to widen our understanding of the universe.
Most reasonable scientists understand that some of these objects will blur the edges—like many moons being more massive and larger than Mercury. Or like the largest super-Earth exoplanets being nearly as massive as Neptune, or the largest gas giant exoplanets being 30-80 times as massive as Jupiter and hence having a surface temperature of ~2000 K. But that is part of science, and possibly even where the most interesting work is.
(This uses wasm, meant for a desktop, and will download the asteroid catalog of about 65mb, you have been warned)
https://dahlend.github.io/ketev/#t=2009-07-01.717&sel=2026%2...
This is a big boy.
I don't agree that "its obviously true that it was the right move". Instead of demoting Pluto, they could have promoted Ceres, and let Eris, Haumea, Makemake, etc, be planets too. Under that proposal, they could even have the category of "dwarf planets" encompassing them, it is just that "dwarf planet" would be a subcategory of planets rather than a disjoint category from them.
Pretty rad.
AI hype organizations like Latent Space (+ others) ban people who disrespect any aspect of AI today, and articles like this get buried.
And then I pull up HN / LinkedIn / X / Substack where everyone is telling me that this is the new AI economy and the tech can't possibly flop, so I figure that the crash must be coming because there's too much optimism.
I go to sleep. Then I wake up and read...
It’s only weird if you come from languages like Java where “primitives” are distinct from “objects”, which is a self-imposed limitation the language imposes.
foo.bar().baz[expensive()].qux.toggle();Long time follower of this. Very grateful it's been implemented.
Any type (like in TypeScript; Rust doesn't have one) is a set of every possible value; a supertype of all types. Every expression regardless of type can be assigned to a variable of type any.
The opposite is the never type, which is an empty set; no value is of type never. Per set theory, an empty set is a subset of every set, which makes the never type a subtype of all types. An expression of type never can be assigned to every variable regardless of type.
If never type has no values, then how can an expression have never type? Easy: it must never evaluate. Infinite loops, unconditional panics/exceptions, process exit/abort, etc. Also, generics - you can have Vec<!> that's always empty, Result<T,!> that can never be an error, etc.
That said, this release includes some incremental improvements as well as the unused-dependency lint to help reduce wasted work within the ecosystem.
Some changes that have been or are about to be merged for upcoming releases
- We will be reducing the debug info generated for the `dev` profile which reduces build-dir size and speeds up builds. We've split off a `debug` profile for when that is needed.
- Cargo will be deleting the directory rustc writes before each invocation, avoiding the leaking of debug files when split-debug is enabled (default on Mac)
- rlibs for workspace members will no longer be copied from build-dir to target-dir which takes up extra space on systems without hardlinking or where we skip it (like Mac)
- redundant information in rlibs and rmeta's is going to be de-duplicated
Examples of further out changes
- This release includes the foundational work to add a build-dir GC
- Cross-workspace de-duplicating of build-dirs
- Cross-workspace build-dir cache
- Early cut off
And the large target directories is also being worked on as well.
Swift handles this through the equivalent of Rust's `dyn`. We could have a stable ABI that's based entirely on `dyn`. Such an ABI would have some limitations, since there are traits `dyn` can't represent (anything that's generic over something other than the type that implements the trait, which is quite common), though we're working on some of those.
Essentially, you wouldn't be able to have a `Vec<T>` in your ABI without either pinning down T or using something like `Vec<Box<dyn Trait>>`. Which is possible, but kind of awful.
We're still planning to ship a stable ABI at some point, but that gives a flavor for some of the kinds of problems we'd need to solve in order for that stable ABI to be useful.
(C++ typically handles generics by having types defined in header files, and those get compiled with the program consuming the library, which is effectively the same situation Rust is in. You could export symbols from a shared library for specific instantiation of a template, but people rarely do that.)
- bool::toggle() https://doc.rust-lang.org/stable/std/primitive.bool.html#met...
- Box::map() https://doc.rust-lang.org/stable/std/boxed/struct.Box.html#m...
- str::trim_prefix(), str::trim_suffix, Path::trim_prefix()
> warning: unused dependency `tracing-subscriber` --> Cargo.toml:23:1 | 23 | tracing-subscriber.workspace = true | ^^^^^^^^^^^^^^^^^^ | = note: `cargo::unused_dependencies` is set to `warn` by default help: consider removing the dependency on `tracing-subscriber`
Heck yes
- Stabilize never type
- Stabilize custom allocators
- Introduced minimum publish age configurations, also known as dependency cooldowns. When the relevant config is set, the resolver ignores versions published more recently than that, unless they are already in Cargo.lock. (RFC 3923) docs #17335
- cargo-install: now uses the existing lockfile in the .crate tarball by default. --locked in cargo install becomes a no-op. #17388
Your claim that supply is effectively fixed also seems false. Transmission investment driven by new demand grew 17x in five years. The problem is that capex deployed to serve data centers is being paid for by households, not that it cannot be deployed.
Real-world data also cuts against you. New York and New England are among the hardest places in the country to build energy infrastructure, yet the study found no clear price increase there from data centers.
The bill you cited would fix much of this, but it would do so via provisions forcing data centers to pay for the costs they cause, instead of forcing households to subsidize them.
If your point is “it’s policy,” fine. But the policy causing this is cost allocation, not permitting.
Your comment doesn’t engage with the article’s central finding at all: who pays.
Congress could fix this problem tomorrow. And in fact there is a bipartisan bill (BAAJA) that would go a long ways (although not all the way) towards fixing it. Now yes even if we fix these policies there would be a time lag on new generation being connected, but there is no reason to leave this problem in place.
Both are solvable with politicians who can be held accountable. But, in many locations, they aren't accountable and can only be trusted to do what's in the interest of their largest donors.
That's all ignoring the various other externalities... aesthetics, noise, pollution.
But, 100%, if a county/town has an underutilized industrial zone, they should build data centers there. But, it also appears data center developers prefer undeveloped plots. Again, also solvable, but somebody needs to do so.
How would we learn about status in case of auth-adjacent incident?
Edit: of course there's the gray area of scrapers, game downloaders etc... but those are not emulators.
Probably not worth losing lifelong customers and PR over piracy concerns.
Anyway, the team behind Yuzu is still at large and the development will continue elsewhere, from aside, like in older days of emulation. Not everything has to default to GitHub, and not everything can.
• Cupcake (Android 1.5)
• Donut (Android 1.6)
• Eclair (Android 2.0–2.1)
• Froyo (Android 2.2)
• Gingerbread (Android 2.3)
• Honeycomb (Android 3.0–3.2)
• Ice Cream Sandwich (Android 4.0)
• Jelly Bean (Android 4.1–4.3)
• KitKat (Android 4.4)
• Lollipop (Android 5.0–5.1)
• Marshmallow (Android 6.0)
• Nougat (Android 7.0–7.1)
• Oreo (Android 8.0–8.1)
• Pie (Android 9)
I want to go all in the Gemini ecosystem but several things keep me going back to OpenAI:
* the Gemini app is glitchy on iPhone (my chats frequently fail or are lost if I background the app)
* the Work equivalent (I guess task?) is not as powerful. I often have ChatGPT Work to arbitrary computation and even install ML on the short-lived VM.
* there is no simple integration between antigravity and the Gemini app. I want a way to continue working away from my server while using the familiar app. No I don’t want VNC.
* ChatGPT sites are cool for putting up proof of concepts with zero effort
They botched it.
A haiku is a short poem. A sonnet is a longer work. An opus is a great work. Mythos is a connected body of works.
As with Open AI, Luna --> Terra --> Sol --> Astra from smallest to largest. Very straightforward.
With Google it's totally incoherent. Hydrogen --> Argon --> Carbon!?
Hydrogen is small, but it's the most abundant element in the universe by far. Argon is >3x larger than carbon by molecular mass, though carbon is more abundant. It's impossible to determine what's supposed to be smaller and what's supposed to be larger.
Thanks any Googlers who had a hand in this
This model naming scheme works for separate products, but using elements, molecules, etc. for products you're supposed to compare with each other... Not very helpful.
If anything, my intuition would say that Hydrogen is more powerful than Argon... I don't know. Not liking this naming scheme trend.
But it is priced the same as the non-discounted 3.8 Flash, considerably higher than what anyone is paying today for 3.8 Flash or 3.5 Flash Lite or the deprecated 3.0/2.x Gemini models. How does this work? Is Google abandoning the “fast and cheap” tier? That seems like a mistake since plenty of workflows are absolutely price inelastic, meaning when the 3.8 Flash discount ends at EOY, lots of current enterprise Gemini customers would have to take their business to another shop??
In order of sophistication and cost, the models are Hydrogen -> Argon -> Carbon.
What kind of ordering is that?!?
With the introduction of Hydrogen, it's no longer ordered alphabetically. That would be Argon → Carbon → Hydrogen.
And it never was ordered by atomic mass, which is arguably the best ordering if you're going to name your model classes after chemical elements. That would be Hydrogen → Carbon → Argon.
It's also not ordered by prevalence in the universe. That ordering would also be Hydrogen → Carbon → Argon.
It's not ordered by electronegativity, boiling point, melting point, or any other remotely defensible chemical property I can think of.
It's not even ordered by letter count, Scrabble score, or NFPA 704 Flammability Rating, all of which would be Argon → Carbon → Hydrogen.
7:52 PM UTC: We've identified that affected servers are failing automated health checks, which is causing them to become unavailable, and this may be related to a recent change to a component involved in processing those checks. We're testing a configuration change on a subset of affected infrastructure to address this and, if successful, will apply it more broadly. In parallel, we've initiated a restart of all affected servers and are validating whether they're successfully returning to service.
7:30 PM UTC: We're aggressively pursuing all pathways to reduce impact. As a precaution, we've paused all deployments and configuration changes while we review recent changes for any correlation to this event. We've determined that the elevated load isn't driven by an overall increase in traffic, but by a large number of servers being unexpectedly unavailable, which is shifting load onto the remaining infrastructure. We're investigating what's causing these servers to become unavailable while we work to safely return them to service.
7:16 PM UTC: We're aggressively pursuing all pathways to reduce impact. As a precaution, we've paused all deployments and configuration changes while we review recent changes for any correlation to this event. We've also moved traffic for some services off the affected infrastructure to further reduce load, with additional traffic moves in progress.
6:59 PM UTC: We've completed redistributing connection traffic across our worldwide infrastructure, and we're moving additional traffic off the affected infrastructure to further reduce load. We're also restarting affected servers to return them to service. In parallel, we're investigating whether the elevated traffic reflects a genuine increase in requests or a shift in how traffic is being routed across our infrastructure.
Status page is down. Much of our org is halted.
Fingers crossed that Azure isn't impacted (seems fine so far)
(https://www.geekwire.com/2026/microsoft-makes-new-voluntary-...)
The blog post summarized the most important changes.
- Stabilize never type
- Stabilize custom allocators
- Introduced minimum publish age configurations, also known as dependency cooldowns. When the relevant config is set, the resolver ignores versions published more recently than that, unless they are already in Cargo.lock. (RFC 3923) docs #17335
- cargo-install: now uses the existing lockfile in the .crate tarball by default. --locked in cargo install becomes a no-op. #17388
So to get any model working on $OLD_HARDWARE, the hardest part of the task is now sourcing the hardware itself and giving the agent ssh access -- a modern agent will be able to source a compiler, write the implementation, optimize it, test in qemu (or any other simulator) and push it onto the device.
Example prompt: "write an inference engine for qwen3 0.6b in pure assembly that runs on this device. Write a script to attach the weights to the resulting binary. Prefix is supplied as a stream of u32 token IDs on stdin. The engine streams u32 token IDs on stdout."
Now Reached 16B models running on iPad Air 1st gen and xiaomi tv box, both have less than 1GB of usable memory, and reached some decent speeds given the hardware.
I'm open sourcing the effort done, kernels developed and results for everyone to re-produce and push more, also the modified models are on hugging face : https://huggingface.co/hyperionaidev
I want a more challenging device, what do you think?
I don't agree. Writing introduces friction that forces you to think and reflect. If you let AI do that part, you are not only skipping the writing, but more importantly, the thinking.
I noticed that the quality of email text has gone up around my work these days. everyone is sounding like a master of drafting emails. Ofcourse Ai has a role in it. But real cool thing about this is, this is pulling up people's thinking to have a structure, minimalism and sequencing of expression. Same as how textile industry has enabled people to have an style and expression in clothing.
That's not to say there isn't value in formal correctness, but the important part is identifying when it's useful and when its more trouble than its worth...
Marketing seems to have dropped the ball on that. I doubt anyone's picking it for the +2% higher points, but being able to answer based on a PDF or call is groundbreaking.
Jev is still faster and cheaper for many of the common text uses.
> how do these models calculate probabilities?
The short answer is: more and better data.
The long answer is: probably the same way any other classifier does. Take your ground truth probabilities and your predicted probabilities. The gap between those is called cross entropy, which is used to push the model towards the correct probabilities.
But maybe this is one of those things that's only obvious in retrospect? I'm sure people were trying something similar years ago with grammar-constrained outputs (for example to only output a simple json structure), but you can't get probabilities out of the model that way. Which makes me wonder: how do these models calculate probabilities?
I guess, once you have been subjected to the infinite horrible examples set by the current president of the USA, this seems like just another day at the office.
I prefer a world where character, integrity and honesty still matter.