OpenWrt One – Open Hardware Router 194p 92c

(no text extracted)

peddling-brink This is the official shop page afaict: https://www.bpi-shop.com/products/banana-pi-openwrt-one-rout...

naturalmovement This thing has no practical purpose. The whole point of OpenWRT is to run it on cheap commodity hardware. This ticks none of those boxes. It has two Ethernet ports, no switch. WHY? Inexplicably can be powered via PoE, makes no sense if its purpose is to hang off your ISP's gateway (which almost certainly lacks PoE supply). PoE feature will never be used. You're not attaching this monstrosity to the ceiling. It's utterly gigantic due to inefficient PCB layout. Why is right to repair important for a throwaway router? Given what will usually fail are the hard to source ASICs. By the time it…

williadc I switched from a Google Wifi to this and found it to be just as stable, but with better range/signal strength, and easier to apply the parental controls I want.

drdaeman Just two Ethernet ports (1+2.5GbE), and it’s dual-band (no 6GHz)… I’m not sure who’s the target audience or what’s the use case.

kennywinker $106usd or $84usd without a case and antennas. That’s a solid price. Wish it had more than 1gb ram - goddamn datacenters.

CoMaps – FOSS Offline Maps 91p 15c

Discover more of your journey - Powered by the community

Plan and navigate your trip abroad with just GPS, no need for mobile data. Search waypoints while on distant hiking trails or bike paths.

The app is designed with privacy in mind - does not identify people, does not track you, and does not collect any information. CoMaps was also audited by Exodus.

Efficiently uses the battery, doesn’t drain your battery like other navigation apps.

People like you are helping build the app by adding locations to OpenStreetMap, giving feedback on features, and contributing code on Codeberg to create great maps together. The project is a fork of Organic Maps and Maps.Me, and driven by an open-source community.

Discover your journey, navigate the world with privacy and community at the forefront.

HelloUsername Probably posted because of related recent discussion on OrganicMaps https://news.ycombinator.com/item?id=48794446

ranger_danger Would love to see a Windows desktop version.

Price per 1M tokens is meaningless 60p 28c

It stops being all about the vibes when the API bill hits you. Many companies are now discovering that AI can indeed be pricey. One habit that might be driving up your AI bill is comparing models by $X per 1M tokens. A lower number should mean lower costs, right? Well, not really.

Each frontier lab has its own tokenizer, which determines how many tokens a body of text is split into. For example, all text in this post so far would’ve been split into 160 tokens for gpt-4o, but that same input would cost you 200 tokens for gpt-4 (1106-preview, generated with tiktokenizer.vercel.app). Even within one frontier lab, OpenAI in this case, model pricing per token is incomparable. Comparing numbers between different labs, especially when they’re constantly tweaking proprietary tokenizers, introduces an error that is hard to measure reliably. Anthropic has recently modified its tokenizer, which resulted in Claude splitting the same text into 30% more tokens. Ceteris paribus, this would be equivalent to a rather steep price hike; however, there is another important factor to take into account.

Even if we ignore the influence of the tokenizer, the other important factor is how much one more token is actually worth. I don’t mean the price of the token, but how much you actually achieve with it. If you’re using AI for serious work, chances are that most of your token consumption is spent on “thinking”, which is often hidden or obscured but billed at the same rate as visible output tokens. This technique can greatly improve output quality; however, the length of that so-called “chain of thought” can become the main factor influencing your overall cost of AI usage — and this can vary wildly.

I’ve picked some of the best current AI models from American frontier labs as well as the best offerings from Chinese labs (which are often pitched as almost as good as American models but for 1/x the cost, often x > 10) and put them in a table below. I’ve also included each model’s score in the Artificial Analysis benchmark, which gives AI models tasks to complete. The goal of AA’s researchers was partly to measure model capabilities and partly to measure how much they were billed for each completed task.

Notice that even though GPT-5.5 is nominally more expensive than Claude Opus 4.8, it completes the benchmark at almost half the cost per task compared with Anthropic’s model. GLM-5.2 is much cheaper per token than both GPT (3.57×/5.68×) and Claude (3.57×/6.82×); however, its cost per task is not proportionally lower, suggesting that it’s less token-efficient than frontier models from the West.

One model that perplexes me is Sonnet 5, since it seems to perform worse than Opus 4.8 while also requiring a higher cost per task due to much lower token efficiency. If someone using it could explain to me what the purpose of this model is, I would be glad to listen. (Conspiracy theory: maybe it’s some sort of psy-op by Anthropic to have a lower sticker cost to coax people into using a less token-efficient model that will ultimately raise their bills?)

DeepSeek V4 Pro seems like the strongest cost-efficiency outlier. Although it scores clearly lower on the intelligence benchmark, its cost per task is extremely low. Fable 5 (Mythos with a security muzzle) seems to show a modest improvement with a price hike of more than 3× compared to GPT-5.5.

Overall, I think this table shows that price per million tokens isn’t a meaningful cost indicator. If you don’t consider the actual cost per task, you will make worse model-selection decisions and be left with inferior performance for a higher price.

zeroonetwothree Well, not totally meaningless but certainly can be misleading.

tidbeck Related to this, for our use case, setting thinking to high instead of low made tasks complete faster and cheaper (Gemini 3.0 flash). Other aspects are caching, often at 0.1X cost, where providers really differ in how efficient they are (Anthropic really good, Google not so much) and how chatty a model is (costing output tokens).

shay_ker cost per benchmark task is definitely interesting! i've always wanted cost per prompt, but even that has too much variation.

A global workspace in language models 123p 39c

As you read this sentence, circuits in your brain are adjusting your posture, controlling your breathing, and transforming lines and curves on the screen into recognizable words. Most of this processing is invisible to you. But some of what takes place in your brain you do have access to—an image that pops into your head, or a deliberate plan you make about where to go shopping. Neuroscientists and philosophers sometimes refer to the latter type of brain activity as “consciously accessible,” to distinguish it from all the other processing that goes on unconsciously. This activity has special properties: we can describe it, control it, and use it for deliberate reasoning, in contrast to all the automatic processing that goes on without our awareness.

In a new paper, we present evidence that a similar distinction has emerged in modern language models like Claude. We find that Claude has developed a small collection of internal neural patterns that, compared to all its other internal processing, play a special role.

We call the collection of these patterns the J-space—named after the technique we used to find them, involving a mathematical concept called the Jacobian. Each J-space pattern is linked to a particular word. But when one of these patterns lights up, it doesn’t mean the model is saying that word—just that the word is on its mind. If you've heard of language models having a "scratchpad" or “chain of thought”—text they write to themselves while reasoning—the J-space is something different. It operates silently, in the model’s internal neural activations, allowing the model to think about a concept without writing it down. Notably, the J-space wasn’t designed or programmed by us, but instead emerged on its own during Claude’s training process.

We find that the J-space has a number of unique properties, compared to the rest of Claude's processing:

Claude can report on these representations. If you ask Claude what it's thinking about, it will tell you what’s in the J-space. Non-J-space representations are less reportable.

It can also modulate them on request. If you ask Claude to think about something, or solve a problem silently in its head, it will light up the appropriate patterns in its J-space. By contrast, it has trouble modulating patterns not in the J-space.

Claude uses its J-space for internal reasoning. If you ask Claude to solve a problem that requires multiple steps, the intermediate steps will light up in its J-space, even when it doesn’t say them out loud. These J-space patterns causally mediate its performance in such tasks, despite being smaller in magnitude than other representations.

Representations in the J-space can be used flexibly for many tasks—for example, once “France” has lit up in Claude’s J-space, the model can recall its capital, or its national currency, or the continent it belongs to.

However, despite its important role, the J-space is not involved in most of what a language model does—speaking fluently, recalling simple facts, using correct grammar, etc. In experiments where we prevented Claude from using its J-space, it still interacted normally, but lost its higher-order cognitive functions.

Our experiments were inspired by a prominent theory in neuroscience that was developed to explain how conscious access works: the global workspace theory. This account pictures the brain as a collection of specialist systems that work in parallel, unconsciously, and largely in isolation from one another. A piece of information becomes consciously accessible when it gains entry to a small shared channel, the “workspace,” which is broadcast to other brain systems that can see it and make use of it. Based on our findings, we think the J-space plays a similar “workspace” role in Claude. For example, we find evidence that Claude’s J-space has especially strong connections to the rest of its neural network, allowing it to fulfill this kind of broadcasting role.

None of this tells us whether Claude is conscious in the way people are, or whether it feels anything at all; we’ll come back to that question at the end of the post. But whatever its philosophical significance, the J-space is a practically useful tool for us, as it gives us a way to see what Claude is thinking but not saying. For instance, we’re able to use it to catch Claude privately noticing that it’s being tested, intentionally producing fabricated data, or pursuing a hidden goal that we planted during training. We’ve also developed a technique to…

esafak Without using the term, they are using an information geometric approach.

meatmanek It would be really cool if they could expose this information to customers somehow. Imagine: - having a log of the most prominent J-space tokens during your customer support chatbot's interactions with a user, so you can have more introspection into why a particular outcome happened - being able to detect certain thoughts associated with undesirable behavior (hallucinations, overstepping authority, lying, etc.) and trigger some sort of remediation (e.g. upgrading to a better model, redirecting to a human, forcing tool calls)

eamag Is it scaling up of https://openreview.net/forum?id=w7LU2s14kE with some changes on where this method is applied?

bilsbie I’m confused where in the weights the jspace is.

Linux on the Atari Jaguar. No, really. 61p 6c

What in the tarnation is an Atari Jaguar?

Why in the tarnation Linux of all things?

Misc. config. files if anyone wants to try

What in the tarnation is an Atari Jaguar?

Released in North America in November of 1993, the Atari Jaguar promised to be the new cool kid in the block thanks to it's (Highly debated) 64 bits of pure power.

The console itself ended up being a commercial disaster, even after the release of the CD addon, the Jaguar CD; which managed to sell even less units in a desperate attempt to try and compete with the Sony Playstation and the Sega Saturn.

Why in the tarnation Linux of all things?

Interestingly enough, to this day, Linux has architecture code for the 68000-family of processors. 68040, 68030, 68010... and even the original base 68000 processor. All neatly structured under arch/m68k/.

As a refresher, the Motorola 68000 was a CISC processor with mixed 16-32 bit capabilities (It's usually described as being 32-bit internally due to the register width length and 16-bit because the data bus was 16-bit, so 2-byte transfers at a time).

It had a 24-bit address bus, 2 to the power 24; thereabout a maximum of 16MBs of addressable memory.

It was released on 1979 to compete with the soon-to-be-released 16-bit CPUs of the era.

Overall, it got lots of traction commercially; it ended up being incoporated in lots of popular commercial hardware. Biggest contenders are the original Macintosh (And Apple Lisa), the Commodore Amiga series of computers (With varying generations of the 68000), the Sega Genesis/Megadrive, the Neo-Geo AES, Plexus workstations... and the Jaguar.

Let's now address the elephant in the room, why Linux.

Doing a bringup for a new 68000-based Linux port should be easy... right? Well... you're in for a good time.

As you may know (Or not), there's an extended thought that Linux requires an MMU to run (You know, being able to use virtual memory is a good thing if we want to run software w/o having day-long headaches).

Technically you are kind of right, but there's uClinux; which precisely makes this feasible. At some point it stopped being a downstream fork of Linux to become part of it; thankfully it's baked in for m68k (And well, the flat memory model and the rest of requirements you can imagine that an MMU-less system has).

Okay so, we enable all the required configuration flags on the Linux menuconfig (To basically tell it not to use an MMU and use the Flat Memory model), we compile and it should run right? Well yes... but no.

The Jaguar has 2 megabytes of RAM (Mapped at 0x000000), (up to) 6 megabytes of ROM (The cartridge, basically) mapped at 0x80000 and 2 custom ICs (Tom & Jerry, a "GPU" and a DSP) which are also memory-mapped.

The main hurdle we have is the memory footprint; while it's true that it's a good amount, it's still not infinite (We're talking megabytes, not gigabytes).

We basically need to find a way to try and optimize our RAM usage. We can do the easy things first, removing features from the kernel, disabling debug... the usual suspects; but we'll still probably have the issue where we can't load the kernel to RAM w/o going OOM (Heck, we're not even considering an initramfs at this point, which are also bulky).

Thankfully, Linux is smart enough to let us "split" the kernel in 2 separate memory regions. One can take advantage of the fact that we can store the read-only sections of it (Think, .rodata or .text) on the cartidge (The ROM) and the dynamic sections (.data or .bss) in RAM (Think, XIP; eXecute-In-Place).

Fortunately for us, it's a matter of telling it where we have the RAM, and where we have the ROM and it (Liunx) handles the relocations for us.

Cool, we can now fit Linux on the Jaguar and it should be able to execute... right? Well sure, you can try specifying the base of the kernel (Remember, non-PIC code) and loading it there and just doing a jmp $80000... but how do we know what is happening under the hood?

In the context of booting Linux, we need 2 things (At least, when doing the bringup):

Any type of output so we can see kernel messages

Any way to tick the system (A timer, basically)

The first requisite usually involves ye' good ol' UART. The Jaguar's DSP (Jerry) has TXD & RXD pins that (At least for booting Linux) can be repurposed to do serial output (If we ignore anything that has to do with sound). Writing a small console driver that bitbangs the pins is enough to start seeing some earlyprintk messages.

The second one, is a bit tougher. But we can basically use any of the 2 timers…

cakehonolulu This is a deep dive on what is necessary to get Linux on the 68000-based Atari Jaguar. No specialized hardware/flash carts. All runs within the original hardware vision (2 megabytes of RAM) and gets to a Busybox shell. Linux repository with the changes: https://github.com/cakehonolulu/linux_jag

basilikum How long does it take to boot?

Python 3.14 compiled to metal – no interpreter 39p 14c

GitHub Copilot appDirect agents from issue to merge

GitHub Advanced SecurityFind and fix vulnerabilities

Code securitySecure your code as you build

Secret protectionStop leaks before they start

GitHub SponsorsFund open source developers

Enterprise platformAI-powered developer platform

GitHub Advanced SecurityEnterprise-grade security features

Copilot for BusinessEnterprise-grade AI features

Premium SupportEnterprise-grade 24/7 support

We read every piece of feedback, and take your input very seriously.

Use saved searches to filter your results more quickly

To see all available qualifiers, see our documentation.

Notifications You must be signed in to change notification settings

pon is a JIT & AoT native compiler and runtime for Python 3.14, written in Rust. There is no interpreter and no bytecode: every module is parsed with the ruff parser, lowered to one shared IR, and compiled to machine code through Cranelift — either just-in-time inside the process (pon run) or ahead-of-time into a standalone native executable (pon build). Memory is managed by a Green Tea garbage collector instead of reference counting, and correctness is enforced by a byte-exact differential harness against CPython v3.14.0.

The end goal is the bun/v8 of Python: a runtime that passes the CPython test suite, runs a multi-tier JIT well past CPython, ships single-binary executables, and includes batteries (package manager, tooling) out of the box. The project is under heavy active development — see Status for what is true today versus where it is going.

# JIT: parse → IR → Cranelift → run, in-process printf 'def add(a, b):\n return a + b\n\nprint("hello, world")\nprint(add(2, 3))\n' > hello.py cargo run -p pon -- run hello.py # AoT: same IR through cranelift-object, linked into a native executable cargo run -p pon -- build hello.py -o hello ./hello

Both paths print the same bytes CPython would. That property is not aspirational — it is the exit gate of the conformance suite (see Conformance).

pon run <file> [args] pon build <file> -o <out> [--allow-dynamic] [--opt] [--target <triple>] pon repl pon -c 'print(40 + 2)' pon - < script.py

One IR, two backends, one runtime ABI. Every tier — baseline JIT, optimizing JIT, and AoT — lowers the same IR and calls the same pon_* helper functions:

source.py │ ruff parser (pinned 0.14.0, PythonVersion::PY314) AST ──> PON IR (pon-ir, one IR for every tier) │ ├── pon run: pon-codegen ──> cranelift-jit ──> native code in process │ tier-0 baseline (all boxed) │ tier-1 typed: inline caches, OSR, background compile │ └── pon build: pon-codegen ──> cranelift-object ──> object file ──> linked executable │ pon-runtime (object model, builtins, NULL-sentinel pon_* ABI) pon-gc (Green Tea garbage collector)

Object model: CPython's heap object layout minus the refcount header. Errors cross the ABI as NULL sentinels, not unwinding. Integers are arbitrary-precision (num-bigint behind PyLong); a tagged small-int fast path is landing in the typed tier.

Tiering: tier-0 compiles everything boxed with no type feedback and is the correctness baseline (PON_TIER0_ONLY=1 forces it). Runtime helpers feed FeedbackCell type profiles from the first execution; hot functions recompile on a background thread and running loops enter the optimized code via on-stack replacement.

GC: the Green Tea collector owns all Python objects. Tier-0 uses conservative stack scanning with a register-flush trampoline at safepoints; the typed tier upgrades to precise Cranelift user stack maps.

All dependencies are declared once in the root Cargo.toml under [workspace.dependencies]; member crates only inherit (see AGENTS.md).

The correctness contract is differential: a corpus module passes only if pon produces byte-identical output to CPython v3.14.0 (TZ=UTC, PYTHONHASHSEED=0). Passing sets are ratcheted into committed floor files, and CI fails on any regression below the floor.

The standing gate is scripts/gate.sh; only its output counts as a gate claim:

bash scripts/gate.sh fast # build + workspace tests + conformance floor + AoT floor + ft-stress bash scripts/gate.sh full # + cpython-full, bench, tier0-only diff, fuzz, package-manager E2E # Individual meters cargo run -q -p pon-conformance -- --suite cpython --check-floor cargo run -q -p pon-conformance -- --mode aot --suite cpython-aot-subset --check-floor cargo run -q -p pon-conformance -- --suite cpython --modules <file.py> # one module, differential cargo run -q -p pon-conformance -- --suite fuzz --seed 42 --count 200 --jobs 8

Corpu…

ubercore I hate to be that guy, but... one week old project, clear signs of vibing. I will be shocked if the remaining work listed (cpython test suite) proceeds in any reasonable timeline. This is a pretty hard problem to just solve in a week. EDIT: and man, these kind of comments LLM created comments are really starting to grind my gears as my job slowly turns into reviewing LLM PRs: > Known gaps at the language level are burned down through the ratcheted floors above — the committed floor files, not this README, are the authoritative compatibility baseline.

westurner How does performance compare to RustPython compiled in a similar way?

AMD Ryzen AI Halo – $4k AI Dev Kit 225p 163c

Performance with Increasing Context - Agentic Simulation

“Best Known Configurations”(BKC) - Batteries Included

Running and Serving LLMs with LM Studio, Getting Started with Lemonade, & Local LLM Coding with VSCode and Qwen3-Coder

Getting Started- and Fine-Tuning LLMs on PyTorch

Appendix A: USB-C Power Delivery and Negotiation

AI Dev Kit, Batteries Included - AMD Ryzen AI Halo

The AMD Ryzen AI Halo is a truly mini-PC built around the Zen 5 AMD Ryzen AI Max+ 395 processor(16 core, 32 thread) that streamlines learning AI development with ROCm or AMD hardware. The Max+ 395 processor is equipped with AMD Radeon 8060S integrated graphics which will be doing most of the heavy lifting, and an NPU which historically doesn't do much, but we were finally able to use.

It comes in a single hardware configuration with a removable 2 TB M.2 SSD and 128 GB of unified LPDDR5x-8000 memory capable of 256 GB/s bandwidth. 2 TB is a good amount of storage to hoard local models, and 128 GB is certainly enough memory to load a couple reasonably sized models into memory while reserving some space for system operation.

Up to 16-core Zen 5 AMD Ryzen AI Max+ 395

Integrated 40 RDNA 3.5 Compute Unit AMD Radeon 8060S

Up to Integrated 40 RDNA 3.5 Compute Unit AMD Radeon 8060S

The AI Halo can be purchased for $3,999.99 USD in a single hardware configuration, preloaded with either Windows 11 Pro or Linux. You are able to load your own OS on the system once you have it, but as far as we know AMD won’t be making the ‘factory’ Linux and Windows installs(packaged drivers, programs, and models) available.

AMD has sent us the Linux version of the Halo which is running a custom AMD Linux distribution based on Debian 13.4.

> hostnamectl ... Operating System: AMD Ryzen AI Developer Platform 1 (rex) Kernel: Linux 6.18.35+rex+2-amd64 Architecture: x86-64 Hardware Vendor: Advanced Micro Devices, Inc. _AMD_ Hardware Model: RAH-001 ...

Check out the interactive CT scan of the Ryzen AI Halo below!

Despite the marketing images presenting it as the size of a datacentre, the Halo is an incredibly small box with only a square 15 cm(6 in) footprint and at less than 5 cm(2 in) tall. It weighs 1.2 kg, but if you’re planning on putting this in your backpack then also consider the required 240 W power brick.

The power button and all of the ports are on the back face of the chassis: four USB 3.2 Type-C ports, an HDMI 2.1 port, and a 10 GbE ethernet port. Besides the connectivity on the rear it features Wi-Fi 7 and Bluetooth 5.4. The USB Type-C port closest to the power button is dedicated to USB-C Power Delivery(PD) power input.

There aren’t any clear affordances for stacking them, but the corner feet and air intakes on all sides should make it viable if you need Windows and Linux, or if you want to cluster them.

The Halo contains two blower fans to draw air in through the top and sides of the case which is then blown through the heat sink and out the back. This box usually sits quietly, but can ramp up the fans to dissipate the 120 W TDP of the processor inside.

The best feature is the white ring of light around the bottom of the case.(pulsing blue when asleep) It doesn’t cast much light, and it can be turned off, but it gives it a nice look without being gaudy.

Being a tightly integrated mini PC there isn’t too much to see inside but you only have to remove four bolts beneath the removable magnetic feet to lift off the bottom cover.

Inside the Bottom Cover of the Ryzen AI Halo

The removable M.2 2280 SSD is easily accessible with no further dissection. Removing the top shell to expose the compute core only requires addressing a few more connections.

Inside the Bottom Cover of the Ryzen AI Halo - Annotated

The core can be pulled out but there isn’t much else to be done. The bottom metal plate visible when first removing the case bottom is removable with four bolts, but we didn’t remove it so as to not mess with the thermal compound underneath.

The AMD Ryzen AI Max+ 395(Strix Halo) processor has been available since Spring 2025 and the Halo doesn’t offer anything new on that front. The 2 TB SSD and 128 GB of memory are as expected, but it’s all been seen before in other hardware like the Framework Desktop, Beelink GTR9 Pro, X+ Rival, and ACEMAGIC M1A PRO. We’ve run some benchmarks to show that it is capable of what is expected, but the main focus of this product is the ‘batteries included’ software which is covered in the next section.

For previous “AI” specific hardware we’ve used MLPerf and Procyon…

kamranjon In case it saves anyone some time (from the article): "The AMD Ryzen AI Max+ 395(Strix Halo) processor has been available since Spring 2025 and the Halo doesn’t offer anything new on that front." It has the same 256 GB/s memory bandwidth limit as every board previously, not sure why this is even being released right now as if it's some new fangled thing - you can go get a Framework Desktop for roughly the same price or a GMKtec EVO-X2 for a bit cheaper.

htrp Does this have the same memory bandwidth problems as the spark?

alexdns Bosgame is $2799 does the same thing if you plan to run only 1 of them

glimshe How much are we going to pay for "AI kits" once the DRAM shortage is over? Will we be able to run a local model equivalent to the current AI frontier in sub $1000 hardware, even if dedicated, in 5 years?

robotswantdata Was “only” $2k in its previous form but even in this updated box the mem bandwidth is woefully inadequate. There’s a few models with space for a dedicated GPU for hybrid inference but imo not worth it. Save your money for a Xeon or EPYC build

syntaxing I have another strix halo that I got for half the price (before this price increase world wide). AMD making lemonade is one of the best reasons to get a strix halo. Lemonade + qwen3.6 35B MTP @ Q8_0 + anythingLLM (in docker) replaced 90%+ of my AI usage. And it’s fully local! Setting everything up took less than 3 hours total, including installing the OS https://lemonade-server.ai/

lhl The one thing that's new/worth pointing out are the https://developer.amd.com/playbooks/ ( https://github.com/amd/playbooks ) - this is AMD's answer to Nvidia's playbooks ( https://build.nvidia.com/spark / https://github.com/NVIDIA/dgx-spark-playbooks ) - I think it's great that they're actually taking this more seriously. Hardware is the exact same as what used to be available for $2K last year (and is still $1K cheaper from Chinese OEMs). LTT Lab's LLM testing is getting more sophisticated, which is great - I think it's worth noting that ROCm/Vulkan versions and llama.cpp build versions are…

Resetting Xbox 342p 279c

This message was just sent to Team XBOX employees globally.

We are beginning the most significant restructure in XBOX history. After careful consideration, I’ve made the difficult decision to reduce our team by approximately 3,200 throughout FY27. This will include approximately 1,600 role eliminations today, and in addition, four studios will leave XBOX to new management. I recognize that a year-long restructuring creates additional challenges. Unfortunately, it is not possible to make all the necessary changes in a single day, and I wanted to be direct about the scale.

I know this is painful. These changes will directly affect people who have poured their creativity into building XBOX. Many joined us through acquisitions, while others were recruited here, or sought us out because they loved this industry and loved XBOX. Today’s decisions do not reflect their talent or dedication.

Our business today is not healthy. We are operating at margins that are 3-10x lower than comparable platform and publishing businesses. We entered Gen 9 with a smaller install base and a higher cost structure. To grow, we bet on Game Pass, multi-platform, and a broader portfolio of content. While those businesses have created meaningful value, they did not grow at the pace we expected. As that happened, our core business weakened, and we added more teams, more investment, and more time, hoping for a better outcome. And now the industry is facing the most severe hardware crisis in its history. We must reset XBOX.

First, we will reset our content portfolio.

Since 2018, we have aggressively expanded our studio portfolio while the number of games created each month across the industry now outpaces the last ten years combined. We now find ourselves competing not only with the largest publishers, but also with smaller independent studios. It is neither possible nor desirable to own every great independent studio. We have also learned that we are not the best home for every type of studio; in a typical year, we lost 64 cents for every dollar we invested. As we reset XBOX, we will help independent creators succeed by providing open development tools and audiences to realize their vision.

Compulsion Games and Double Fine Productions will return to management and transition to independent studios with their IP, catalog, and runway for their next games. Ninja Theory and Undead Labs have entered terms to join new ownership with funding to complete and grow Senua and State of Decay 3. In France, Arkane’s management is beginning required consultation with its Works Council to review potential strategic options.

We are also making reductions across other units, and in some cases, shifting investment to focus on higher priority projects. These changes vary in size across Activision, Bethesda/ZeniMax, Blizzard, King, Mojang, and XBOX Game Studios. None of our first party publicly announced games or projects are being cancelled as part of these reductions.

In addition, Mojang and King will now report directly to me. These two studios have increasingly become platforms and are our largest by monthly active players. They bring critical geographic, demographic, and differentiation to XBOX.

We know that great technology gets better when it gets simpler, not bigger. Today, in some parts of the company, work passes through as many as 14 layers of management. Our platform teams are 40% larger than they were at the start of this generation, even as our player base and playtime have declined. That complexity has slowed decisions, blurred accountability, and made it harder to deliver for players. As we reset XBOX, we will simplify.

We will reduce management layers to no more than 5, and where possible, 3. We will deliver success through a flatter organization that is built around makers (individual contributors focused on building), player-coaches (leaders who remain deeply involved in the work while developing their teams), and directly responsible individuals (DRIs) who own key decisions and outcomes. And we will streamline how we work across our tools, with a cleaner code base, shared services, and 50% reduced vendor spend.

As XBOX grew our headcount, we became more fragmented. Teams, studios, and functions often operate independently, and it became harder to work towards a shared goal, make the right tradeoffs, and get things done.

For the first time, we are establishing a Chief Operating Officer with end-to-end P&L responsibility across content, hardware, platform, and…

ChrisArchitect Related: Microsoft cuts 4,800 Jobs, Half from Xbox division https://news.ycombinator.com/item?id=48804401

clonedhuman Continuing the long trend of major tech companies making everything they touch worse.

ChocolateGod I think some of these game studios got so content with Microsoft constantly paying that they forgot to make games that would actually sell. South of Midnight took 7 years to make and cost $100 million to make... yet sold hardly any copies and I'm not even sure who they were trying to make it for. Meanwhile you have studios like Sandfall and Warhorse pumping out games on a fraction of the budget that ship millions (and imho, make better games).

1970-01-01 They will just continue smash thru exactly what is killing them because they do not know how to reset. More micro transactions, Halo 14-39, games launching before they're ready, price increases, etc. All of that looks good on paper, so they will take no action against. The XBOX is hitting icebergs, and instead of slowing down, they will just call for more speed.

tangenter Game Pass has caused a lot of direct sales losses to game developers in favor of Microsoft trying to find a Netflix-like cash cow for itself. The numbers never added up, but it is not a surprise everyone nodded and went along with it. I wonder what the career repercussions would be for speaking up - but it doesn’t matter because they are getting fired anyway. Call of Duty alone lost $300 million: https://arstechnica.com/gaming/2026/04/microsofts-game-pass-... I look forward to the source code leaks.

dcrazy Any details about the studio spin-outs? The rumors were that Double Fine etc. would be closed, but all we know now is that some of them are being sold to management and others are being sold to other investors. Nothing about any commensurate restructurings.

speak_plainly At some point, the games industry decided it wanted to be interactive Hollywood, and the consequences are entirely predictable. Meanwhile, Nintendo just quietly shipped 3.8 million units of Tomodachi Life in two weeks, and 4 million of Pokopia in five. They're making actual games. Sony's obsession with prestige cinematic bloat, like Xbox, has also put them in a slow-motion death spiral that's going to become painfully obvious in a few years.

athorax It is neither possible nor desirable to own every great independent studio. We have also learned that we are not the best home for every type of studio This is shockingly self-aware for microsoft

hbn It was pretty rich seeing armchair video game industry analysts act like the new CEO was gonna usher in a new age for Microsoft's gaming division because she got to announce the updated logo and some games that would have obviously been in development long before she became CEO. Microsoft is never going to figure out gaming. It's more art than engineering and they can barely manage the engineering with all the intervention from marketing and HR in their products. To me it's mostly unfortunate that this has left PlayStation with no direct competition because they've noticed and leaned into the…

Using precision editing to study human embryo development shows master gene 19p 3c

First use of precision editing to study human embryo development reveals role of master gene

The Animal Welfare and Ethical Review Body

Scientists have, for the first time, used an extremely precise genome editing technique called base editing to study gene function in human embryos. They found that a gene called NANOG is essential for forming the future body from an embryo.

Base editing can precisely change a single nucleotide base pair to another in an entire human genome of around 3 billion base pairs - that’s an incredible feat.

Research led by the University of Cambridge Loke Centre for Trophoblast Research has shown that a genome editing technique can be used to alter a single gene in human embryonic cells, enabling the study of very early human development in unparalleled detail.

The technique, called base editing, is a more precise version of the genome editing technique CRISPR/Cas9. It can change a single nucleotide base pair - the basic building block of DNA - within a human genome of approximately 3 billion base pairs.

Using base editing, the researchers blocked a gene called NANOG in very early-stage human embryos, and found that the cells of the early embryo could not develop into more specialised pluripotent cells called the epiblast - which later form the body.

The results reveal the crucial role of NANOG in the development of human embryos, and helps scientists better understand how human embryos develop in the first few days after an egg is fertilised.

Without NANOG, the cells that later become the placenta and yolk sac - the tissues that support the developing embryo - could still form.

While human embryo base editing has been previously reported, this is the first time that this technique has been used to study gene function in human embryos. The results show that the extreme precision of the technique reduces the likelihood of unintended chromosomal abnormalities, which can occur with another more widely used version of CRISPR/Cas9.

Understanding more about the role of genes required for human development, such as NANOG, could in future help to improve IVF success rates and better understand early pregnancy loss.

Base editing could also potentially be used in future to edit specific genes for debilitating inherited conditions - like cystic fibrosis and Huntington’s disease - in human embryos to prevent the conditions being passed on to future generations. However, this would not be legally permissible in the UK at present. Before any future clinical use, extensive safety testing, further development of the technique, and broad public debate and support would be required.

“Base editing represents a significant advance on conventional CRISPR/Cas9 because it carries a far lower risk of causing unintended chromosome errors. Base editing can precisely change a single nucleotide base pair to another in an entire human genome of around 3 billion base pairs - that’s an incredible feat,” said Professor Kathy Niakan at the University of Cambridge Loke Centre for Trophoblast Research, who led the study.

She added: “Our results indicate that the NANOG gene is critical for the development of pluripotent cells, the building blocks that are fundamentally important to human development.”

Pluripotent cells can develop into any other type of cell in the body and are widely used in biomedical research, from drug testing to disease modelling. Human embryonic stem cells, which are pluripotent, arise in a part of the developing embryo that has high levels of NANOG activation. This has caused scientists to suspect that NANOG plays an important role in their creation.

“The precision of base editing is a major step from the previous generation of genome editing techniques. This allows us to study early human development with greater confidence,” said Dr Oliver Bower, a researcher at the University of Cambridge’s Loke Centre for Trophoblast Research and first author of the study.

He added: “By pinpointing how genes like NANOG control the development of pluripotent cells, we can make stem-cell systems for biomedical research more predictable and reliable.”

Enlarge Development of base edited human embryos in first week after fertilisation

Human development does not always follow the mouse blueprint

Decades of animal research, particularly in mice, were essential for identifying NANOG as a gene likely to play a major role in early development. But this study shows that NANOG does not function identically in human and mouse embryos.

In previous…

ape4 I wonder what the regulations are for this sort of work

Stealth robotics startup (YC S26) is hiring principal engineers (Palo Alto) 0p 0c

We are a new robotics company based in Palo Alto, California, building wearable robotic devices. Our first product is a device that reduces the physical cost of carrying heavy loads over distance, allowing people to move farther, carry more, and arrive less fatigued. The company is early but already operating in the field, with hardware on real users and a direct feedback loop in the most demanding conditions imaginable. We are backed by Y Combinator and a group of leading deep-technology investors. We recently moved to the Bay Area to grow the team and are hiring three principal engineers (we're in stealth for the time being, but hopefully not for too much longer!). $200-250K + 1-2% equity + full benefits, onsite in Palo Alto. These senior roles each own their domain with full authority, no inherited design decisions, and no one between them and the problem. What you build will go on real users within weeks. - Mechanical (Principal Engineer): own the mechanical platform, including structures, joints, actuator integration, ruggedness, and field serviceability. You have designed hardware for robots, humanoids, or other high-performance actuated systems. - Firmware (Principal Engineer): own the embedded stack, including real-time BLDC motor control, sensor integration (IMUs, encoders, current sensing), OTA, and debug tooling. You have written real-time firmware for physical robots or autonomous systems. - Software (Principal Engineer): own the software infrastructure, including device-to-cloud telemetry, test and log databases, data pipelines, and user applications. You have built software for hardware fleets and the people who use them. Please send a note to chrome.braindance@gmail.com that shows what you've built. A resume works, but a quick description of a real project is better. Photos, videos, and links are welcome. We look forward to meeting you!

20:53 UTC · cached 5 min · ?n=1–15