Six stories cleared 200 points today, and they split cleanly into two camps: people building the infrastructure that the AI era actually runs on, and people who think the whole thing has gone off the rails. Both camps are right.
Fujitsu Launches Made-in-Japan Next-Generation CPU FUJITSU-MONAKA
473 points · Fujitsu
Fujitsu’s 144-core Armv9.3-A server CPU finally has a ship date. Global sales of the standalone processor start in November 2026, alongside 1U and 2U MONAKA servers in Japan and Europe, with broader shipments from April 2027. The packaging is the interesting part: the compute die is 2nm (TSMC N2P), but the entire last-level cache lives on a separate 5nm die that the core die stacks onto face-to-face via hybrid bonding, with a third 5nm die handling I/O across a silicon interposer. Fujitsu keeps 2nm silicon under 30% of total die area specifically because it lets them get a 2nm part to market without paying 2nm prices for SRAM that barely shrinks.
Two SKUs: a 350W air-cooled part at 2.1 GHz base, and a 500W liquid-cooled part at 2.9 GHz base, both peaking at 3.8 GHz, both carrying all 144 cores. Memory is 12 channels of DDR5-8800; each socket brings 96 lanes of PCIe 6.0 plus CXL 3.0, and the chip can present itself as eight 18-core NUMA nodes, four 36-core nodes, or one flat 144-core domain. Fujitsu credits ultra-low-voltage operation — running the cores around 30% below nominal — for holding 144 cores inside the 350W envelope, and rates the 350W part at 4,355 GFLOPS DGEMM and 69.7 TOPS INT8.
Now the skeptical part. Fujitsu claims 2× AI inference throughput versus “other CPUs,” roughly 50% TCO reduction, and up to 80% lower cooling power — with no named baselines, no configurations, and no independent benchmarks. Those DGEMM and STREAM figures are Fujitsu estimates, not measurements, and won’t mean anything until someone outside the company runs the tests. There’s also a real design regression to name: the cores run two 256-bit SVE2 units, narrower than the 512-bit SVE in A64FX, the chip that powered Fugaku. Fujitsu’s stated reason is core size and cost, which is fair for a data center part, but it’s a step down from the company’s own HPC heritage, not up. And the “made in Japan” label deserves scrutiny: the compute die is fabricated by TSMC in Taiwan, and the sovereign pitch rests on where the chip is assembled and traced (the Kasashima plant), not where the leading-edge silicon is made. That’s meaningful for supply-chain traceability in defense procurement. It is not fab independence, and the press release is happy to let you conflate the two.
By the time MONAKA reaches volume production in 2027, 144 cores will be mid-pack against Graviton5’s 192 cores and Ampere’s roadmap to 512. Its differentiation is the cache-on-a-separate-die stack, memory bandwidth, and CXL pooling — plus NVLink Fusion integration with NVIDIA, which is the real tell about where the industry is going: Japan’s flagship supercomputer plans, FugakuNEXT, put GPUs at the center.
Hister: A Private Search Engine for the Pages You Visit and the Files You Keep
387 points · GitHub
Hister is by asciimoo, the author of SearXNG, and it’s a different animal from that project. Instead of federating other people’s search engines, it builds a personal full-text index of the pages you visit and the local directories you point it at — then lets you query it from a web UI at 127.0.0.1:4433, a TUI, the command line, or an MCP server an AI assistant can call. Browser extensions for Firefox and Chrome push page content to your own server as you browse; you can also import browser history or run its crawler against sites you choose. It’s Go, AGPL-3.0, 4.2k stars, and the quickstart is genuinely a single binary and one command.
Two things make this more than a nice weekend tool. First, the boring-but-hard part is done properly: field filters, phrases, wildcards, negation, result priorities, and multi-user separation on a shared server, with no telemetry and no cloud dependency in the default configuration. Second, the MCP interface is the compelling use case. An agent that can search a full-text index of your own reading history and files is materially different from one searching the open web — it answers from things you actually looked at, with citations you can check, instead of synthesizing a plausible answer.
The privacy claims are accurate but narrower than they sound. Optional semantic search works by sending document text to whatever embeddings endpoint you configure — turn that on and you’ve reintroduced exactly the third-party dependency the README is proud of avoiding elsewhere. There’s also no mention of encryption at rest, and a complete full-text index of everything a person reads is an unusually attractive target. The comparison shopping matters too: Recoll and DocFetcher already do local full-text search, and Khoj occupies adjacent territory. Hister’s edge is browser-first capture plus the MCP surface, not the search engine itself. It isn’t a Google replacement and doesn’t pretend to be — it’s a memory prosthetic for people who read more than they can recall.
How GLM Built Its Own Inference Infrastructure
349 points · Z.ai
Buried under a title about recursive self-improvement is one of the more concrete engineering write-ups of the year. ZhiPu built production inference for GLM-5.3-Flash from scratch on a cluster of more than 100,000 Chinese-made AI accelerators — a scale nobody had deployed on that hardware before. The constraints were brutal: less chip memory and bandwidth than NVIDIA parts, an immature software ecosystem, incomplete kernel support, and documentation that had to be reverse-engineered. The resulting stack is a pile of aggressive memory tricks: intra-node tensor parallelism for linear attention and the LM head, ReplaySSM (trading compute for bandwidth), W8A8 quantization, mixed-precision KV cache across INT8/FP8/BF16, layer splitting, and an Encode-Prefill-Decode disaggregated architecture. Together they bought roughly 3× end-to-end serving throughput, and the company says utilization and per-token cost are now comparable to mainstream GPUs — a claim with no disclosed benchmark behind it, so treat it as a target, not a measurement.
The part worth stealing isn’t the kernel work. It’s the model of what made the agent useful. ZhiPu’s own framing is that an Infra Agent powered by GLM-5.3 did much of the work, and that its effectiveness depended less on the model’s coding ability than on whether the system could produce attributable feedback — local, cheap, objective, and traceable to a specific layer. Their three case studies are excellent and specific. A KDA kernel’s Context Parallelism path was silently computing in TF32 inside tl.dot on FP32 inputs, accumulating error over long contexts; the fix was input_precision="tf32x3", merged upstream into Flash Linear Attention. KV transfer never overlapped with DeepEP dispatch because the intranode paths held the Python GIL — the internode code already released it, with a comment explaining why — so unlocking the GIL took prefill-plus-transfer overhead from over 20% to under 1%. And a decode kernel tiled along the V dimension repeated the same FP32 normalization and gating four times; merging the tiles into one thread block and keeping intermediates register-resident got 1.71× over the previous version. That’s real, cited, upstream-merged work.
What it is not is recursive self-improvement, and the post’s honesty about that is better than its headline. Engineers defined objectives and system boundaries, built the feedback environment, and reviewed anything touching numerical semantics, concurrency, or production risk. The agent proposed, implemented, and tested. The headline numbers — under two weeks from adaptation to production, 3× throughput, 100,000 accelerators — describe a very good CI loop with a model inside it, which is a smaller claim than “the model optimizes the system; the system runs the model.” It’s also the most persuasive argument in today’s batch that the current ceiling on agentic engineering is instrumentation quality, not model capability. Finally, note that the same write-up casually mentions that security partners used the model to find thousands of real vulnerabilities — which is precisely the dynamic the next story is furious about.
One Year of Sponsored Servo Development
337 points · Servo
Servo, the browser engine now under Linux Foundation Europe, spent a year funding one long-time maintainer — Josh Bowman-Matthews — part-time through OpenCollective and GitHub donations. His retrospective is almost aggressively unglamorous, which is the point. In twelve months: 8 new maintainers nominated, 1,150 pull requests reviewed, 114 issues filed specifically for newcomers with a 92% fix rate, new documentation on borrow hazards, experimental features, finding work to do, and diagnosing flaky test failures. He also supported a large rewrite of the JS engine integration that fixed intermittent garbage-collection panics, uncovered broken window.open behavior, and stabilized a pile of intermittent tests — plus helped another contributor get a grant proposal approved.
If you’ve ever wondered what actually keeps a browser engine alive, it’s this. Not clever architecture posts, but review bandwidth, a maintainer pipeline, and someone whose full-time job is making other people’s contributions land. The 92% newcomer fix rate is the standout number: it means the project is actively closing the loop for new contributors instead of letting their issues rot, which is the single most common failure mode in volunteer infrastructure.
The retrospective also leaves the hard questions unasked. There are no WPT pass-rate numbers, no embedding or adoption figures, no total dollars raised, and no statement of whether the role continues. Servo’s pitch is that it’s a lightweight engine for embedding web tech in applications — what that pitch still needs is a shipping product where Servo is inside. One detail is worth flagging for its own sake: among the docs he wrote is the project’s AI policy. Even a browser engine with no AI features now needs a written position on AI contributions, because otherwise it gets flooded.
CCC Invites All Model Citizens to 40C3
313 points · CCC
The Chaos Computer Club has opened the Call for Participation for the 40th Chaos Communication Congress: December 27–30, 2026, at the Hamburg Exhibition halls, under the motto “Model Citizens.” Calls are open for content, art, punk, and music. The move to Hamburg is the year’s biggest operational change — the club acknowledges that after the venue switch there’s finally room to scale up the event, which for a conference run by volunteers and financed by nobody is either an opportunity or a logistics problem, depending on how the next three months go.
The framing is explicitly political, and CCC doesn’t hedge. They argue that solidarity toward a shared future is giving way to an authoritarian model where the stronger impose their views on the weaker, that liberal forces are playing defense rather than building, and that the Congress is itself the counter-model: difference treated as an asset, the unfamiliar as inspiration. The motto does double duty as a jab at AI’s “model citizens,” which is either clever or too cute depending on your tolerance.
Say what you like about the rhetoric — the concrete claim, that attendees actively shape the event rather than consuming it, has held up for thirty-nine years. Congress is the largest non-commercial hacker conference in Europe, with 16,000+ visitors and a lecture program that has seeded a lot of tooling and policy work that later went mainstream. The worth of a congress is measured in what happens after it, and this one has a track record. If you have something to say about building systems that don’t make people property of a platform, the deadline is now, and the venue has room.
Everybody’s Lost Their Minds
249 points · netmeister.org
Jan Schaumann’s post is a rant, and it’s the most-commented story in the set, which tells you something about the mood. The core argument runs like this: every company is “ethics aside”-ing the known harms of AI — copyright laundering, CSAM generation as a product feature for logged-in users, military targeting, water and power consumption that nobody wants near their house — because “tokens” function as casino chips and nobody has to show ROI. Meanwhile, people with no engineering background are pitching industry-changing solutions built in agent-infested homelabs, and prose everywhere has collapsed into LinkedIn-fluencer cadence.
The strongest section is the one that’s actually falsifiable, about the AI vulnerability-hunting programs. Anthropic and OpenAI have been one-upping each other on how dangerous their models are, and everyone who considers themselves an industry leader joined a mysteriously named project or signed an open letter. Those programs consumed dozens of senior security engineers for months, redirecting priorities and forcing thousands of findings into vulnerability management pipelines — many millions of dollars per organization — and produced thousands of vulnerabilities, only a fraction of which were reported upstream. His conclusion: we are not measurably safer, because finding and fixing vulnerabilities was never the bottleneck. Patching is. Getting packages updated is. He’s largely right, and his counterfactual is the right one to ask: that money spent on asset inventories with package-level granularity, automated update infrastructure, forced reboots at 30 days of uptime, and attack surface enumeration would have moved the needle more than another harness.
Then he turns the blade on the industry’s own cognitive processes, and this is where the piece earns its anger: AI finds vulnerabilities, AI writes the patches, AI reviews the resulting pull requests, and the “human in the loop” is a rubber stamp. Every cycle removes a human who understood the subsystem. Complex systems fail in complex ways, debugging is an order of magnitude harder than writing, and debugging distributed systems of components opaque to their own operators is not a hard problem — it’s an impossible one. The post isn’t balanced and doesn’t want to be, and the author cheerfully admits he spends 75% of his working time on AI, which makes this a burnout document as much as an argument. Read it for the bottleneck analysis, discount the tone, and note the uncomfortable overlap with the GLM write-up above: the same companies building the agentic loops he’s describing are publishing posts about how their models now help build their own infrastructure, and treating that as the good news.
The Throughline
Two things are true at once on today’s front page. The plumbing of the AI era is being rebuilt from the silicon up — a Japanese CPU that stacks its cache on a separate die to dodge 2nm economics, a Chinese model stack running production traffic on 100,000 domestic accelerators because it had no other choice, a personal index that puts an agent’s retrieval inside your own machine instead of someone’s API. And nearly everyone doing the building is quietly anxious about it. Fujitsu’s numbers come with unnamed baselines. ZhiPu’s RSI headline collapses into a very good feedback loop with adults in the room. Schaumann can’t find anyone who’s checked the ROI.
The sovereignty theme runs through three stories and keeps half-keeping its promise. MONAKA is “made in Japan” but fabbed in Taiwan. GLM’s accelerator stack is genuine independence from NVIDIA, paired with a per-token-cost claim against NVIDIA that’s asserted rather than demonstrated. Hister is genuinely local, right up until you enable the embeddings endpoint you configured. Each one trades a visible dependency for an invisible one, and the marketing never mentions the second half.
The other connective thread is feedback. Servo’s year shows that what makes an open-source project durable is not talent but review bandwidth, a maintainer pipeline, and someone closing the loop for newcomers — 92% of newcomer issues fixed is the number that matters. ZhiPu’s year shows the same principle applied to agents: the model was useful in proportion to how cheaply and precisely it could be told that a specific change made a specific thing worse. Schaumann’s complaint is that the industry has invested in the opposite — more generation, more automation, less attribution, less understanding. The difference between those two paths isn’t optimism versus pessimism. It’s whether anyone bothered to instrument the feedback loop, and whether a human who understands the system is still in it.