Invisible Watermarks, a $4 Sol, and the IPFS Sunset

Good morning. It's Tuesday, August 25, and we're covering the invisible GUIDs Microsoft embeds in every locally generated Paint and Photos image, OpenAI's second price cut of the quarter (GPT-5.6 Sol is now $4/$20 through at least November), and the end of IPFS as a funded, maintained project. Plus: Thomson Reuters trains its own frontier model for $40 million, Xiaomi's new chip trades blows with Apple's cores, and a 500-comment argument about whether AI reliance is blocking the next generation of experts.

The Paint reverse-engineering thread held 720 points with 318 comments overnight, the Xiaomi benchmark sat at 880 points with 623 comments, and the "expertise collapse" essay pulled 514 points with 505 comments - the day's three biggest developer discussions are about provenance, silicon, and skill. Here is the signal, sourced.

In today's brief:

• Paint and Photos hide a server-issued GUID in the pixels of locally generated AI images, and "local" still means your prompt is moderated by Microsoft first
• OpenAI cuts GPT-5.6 Sol 20 percent on input and 33 percent on output, in place until at least November 21
• Thomson Reuters ships Thomson, a domain-specialized model trained for $40 million, with a small open-weights version on Hugging Face
• Shipyard confirms it stops maintaining Kubo, Helia, and the ipfs.io infrastructure on September 30
• Xiaomi's XRING 03 pairs ARM C1-Ultra cores with 44 MB of cache - and Apple may answer soon

THE BIG ONE

Paint and Photos Embed Server-Issued GUIDs as Invisible Watermarks in Local Images

Xusheng Li's reverse engineering deep-dive (720 points, 318 comments) documents something Microsoft discloses nowhere: Paint and Photos both embed an invisible, server-issued GUID into the pixels of every AI-generated image, even when generation happens entirely on the device. Following the trail from a suspiciously large Watermarker.dll, Li traced the call chain from a local Stable Diffusion run through WmkWriteWatermark, which takes a 16-byte GUID, wraps it with a checksum byte, expands the 18 bytes into 144 bits, and quantizes selected image blocks to carry each bit at least three times - a content-adaptive, SVD-style encoding that changed 193,376 of 262,144 pixels in a synthetic 512-by-512 test image.

The GUID comes from Microsoft's own prompt-moderation endpoint. Before Paint runs its local model, AIServices.dll POSTs the prompt to apsaiservices-...azurefd.net/v1/paint-cocreator/moderate-prompt and gets back a revised prompt, a promptGenerationId, and a watermarkId. The watermarkId becomes the GUID in the pixels, and it is also recorded in the signed C2PA manifest under com.microsoft.invismark.1 as a soft binding, so the content can still be matched to its provenance record after the file-level manifest is stripped. On a Copilot+ PC the NPU generates the image locally, but the loop is never offline: prompt moderation is remote, and the finished result goes back through online provenance signing.

The details that matter for users: the visible-watermark setting does not control the invisible one, and a failed watermark embed turns the whole generation into an error in Paint (Photos logs and continues). Save formats for AI output are restricted to PNG, JPEG, GIF, and Paint's own .paint format - BMP is disallowed precisely because C2PA cannot embed a manifest there. Microsoft's support pages disclose remote content filtering and the C2PA manifest, and the EU AI Act's Article 50 transparency code took effect August 2, 2026, requiring a detectable machine-readable mark. What the docs do not say is that the GUID on your pixels is an identifier issued by a Microsoft moderation server, linkable across generations via lastPromptGenerationId.

Why it matters: "generated locally" now means the pixels are made on your NPU but the provenance trail runs through someone else's server, so treat every AI image from Paint or Photos as carrying a prompt-linked identifier, and remember that AI images are only watermark-free if you strip the pixels alongside the metadata. Anthropic applies the same provenance logic to text output with Claude's text watermark and C2PA system, which we covered when it shipped.

PRICING

OpenAI Cuts GPT-5.6 Sol to $4/$20, 20 and 33 Percent Off, Through at Least November 21

OpenAI's API pricing page now lists GPT-5.6 Sol at $4.00 per million input tokens ($0.40 cached, $5.00 cache writes) and $20.00 output, down from $5.00 and $30.00 at launch - a 20 percent cut on input and 33 percent on output, with the rate explicitly in place until at least November 21, 2026 (325 points, 312 comments). It follows the July 30 round that took Luna down 80 percent and Terra 20 percent, which we covered at the time. The deeper context: OpenAI's own announcement signals this is the new pricing floor, not a sale, and the thread's most-upvoted reading is about pressure, not generosity - the commenters note Anthropic's demand-trough coverage and see a frontier vendor reaching for price to grow volume.

For developers the mechanics matter more than the sentiment. Sol is the tier that runs agent loops and heavy reasoning, and the cut flows through to Codex and ChatGPT Work usage accounting, so subscription quotas stretch further without any plan change. The HN thread clarified one common confusion: "Subscription usage remains unchanged" per OpenAI's communications; this is an API price change, and ChatGPT subscribers do not get extra weekly quota from it directly. Terra and Luna pricing stay where they were ($2/$12 and $0.20/$1.20), which keeps the cost-per-task ladder intact: Luna for bulk extraction, Terra for everyday agent work, Sol for the hard allocation.

Why it matters: the third straight quarter of frontier price cuts is now measurable in the API table itself, so re-check your cost-per-task math for anything that routes heavy reasoning to Sol - 20 percent on input and a third off output changes which workloads justify the frontier tier at all. Our Sol developer guide maps what the tier is good for, and the 2026 pricing matrix keeps the comparison against every other vendor current.

MODELS

Thomson Reuters Trains "Thomson": $40 Million, Open-Source Base, Own Content

Thomson Reuters' launch announcement (118 points, 46 comments) describes Thomson, its first proprietary LLM, as the "different path" argument made by a data company: instead of billions on training runs, it invested $40 million in talent and compute, started from a strong open-source foundation, and applied mid-training and post-training over decades of Westlaw, Practical Law, Checkpoint, and Reuters content, with "hundreds of subject matter experts integrated from the design of training objectives through to the final evaluations." Early evals put Thomson "on par with the latest frontier models" across a range of tasks - while trained on less than 10 percent of the company's content so far, and it debuts inside Tabular Analysis in CoCounsel Legal, with the product staying multi-model by design.

Two details separate this from a corporate PR cycle. First, the economics: Thomson Reuters is publishing a path where a focused model costs 40 million dollars and beats the assumption that frontier capability requires frontier budgets, even if it is domain-limited by construction. Second, sovereignty: the release positions the model around Fiduciary-Grade standards, where customer data is not used to train without consent, models are owned and controlled by the firm, and deployment can happen in sovereign environments. Legal academics who tested it - including a Washington University tax professor who preferred Thomson over ChatGPT and Claude for corporate-tax questions with links to treatises - are quoted in the release, and the "small" version is available as open weights on Hugging Face for academic and non-commercial use.

Why it matters: content owners are now in the model business with an open-weights escape hatch and a measured cost story, which strengthens the argument that proprietary data and domain workflow, not scale alone, is the moat in vertical AI - the domain-expertise thesis we outlined in July, now with a $40 million production example attached.

INFRASTRUCTURE

IPFS at Shipyard Winds Down September 30: Kubo, the Gateways, and the Bootstrap Nodes

The announcement from IPFS Shipyard (346 points, 177 comments) is signed by Cameron Wood and Adin Schmahmann - Schmahmann is one of IPFS's founding engineers, which tells you how far up the stack this cut goes. Protocol Labs is not renewing Shipyard's funding, so its final day of IPFS engineering, maintenance, and infrastructure work is September 30, 2026. After that, no dedicated maintainers remain for the core implementations: Kubo, Helia, Boxo, Rainbow, IPFS Desktop, IPFS Companion, Someguy, the Service Worker Gateway, and IPFS Check. The public infrastructure Shipyard operates - ipfs.io, dweb.link, check.ipfs.network, delegated-ipfs.dev, the bootstrap nodes, and the collaborative clusters including Wikipedia-on-IPFS - will stop being operated, with Protocol Labs, which owns the domains, deciding what happens next. Shipyard's own summary says its free tools served over 75 million monthly active users.

The backstory makes the timing sting. Shipyard's 2025-2026 direction was explicitly an HTTP-native reboot: moving gateways to inbrowser.link, cutting gateway operating costs by around 80 percent while tripling traffic, and arguing for "verifiable websites and downloads directly in the browser." The community's 177 comments mostly land in two camps: mourning the loss of the last funded IPFS stewardship, and noting that the HTTP-native turn had already reduced reliance on the traditional stack. Neither camp changes the practical question for anyone who still fetches things over ipfs.io or pins through Kubo-based tooling: the layers you depend on are about to have no committed upstream.

Why it matters: content-addressed distribution does not die with Shipyard, but the probability that a bug in your IPFS dependency gets fixed, or that a gateway you rely on stays up, just dropped materially - audit your ipfs.io and dweb.link usage now and have a fallback path for September 30.

HARDWARE

Xiaomi's XRING 03: Apple-Class Single-Thread Performance, Stronger Multithread, 44 MB of Cache

Daniel Lemire's benchmark summary on X (880 points, 623 comments) is the day's biggest hardware thread: Xiaomi's new XRING 03 roughly matches Apple's cores on single-threaded work and is meaningfully faster in multithreaded execution. The specs behind it: a 6+4 cluster with two ARM C1-Ultra cores at 4.35 GHz and four C1-Premium cores at 3.68 GHz, a 133-square-millimeter die with 16 MB of SLC on top of roughly 44 MB of total cache - more cache than most laptop CPUs - plus SME2 for matrix and AI acceleration and SVE2 for SIMD. The C1-Ultra is strikingly wide: 21 execution ports, six of them 128-bit SIMD, which is more ports than any current Intel or AMD core can point at.

The thread adds the caveats Lemire's post glosses over. The cores are ARM C1-Ultra designs licensed by ARM and built on TSMC - the same play Qualcomm runs in the Snapdragon line - so the story is less "Xiaomi beat Apple" and more "ARM's first-party core designs have reached Apple parity, and Xiaomi integrated them first with an aggressive cache budget" (thread context). The multithreaded edge comes partly from that cache and partly from sheer port width, which matters most for workloads with lots of independent arithmetic. Lemire's own read is a trend, not a single chip: "We are getting cores that are massively parallel in terms of the number of execution units... this is where all the transistors go," he wrote, adding that Apple may answer with its next processor soon.

Why it matters: if licensed ARM cores hit Apple's single-thread performance while shipping more execution width and cache, the practical center of gravity for on-device inference and parallel workloads keeps shifting to whoever integrates ARM's flagship IP first - and the SME2 math blocks give local-model runtimes a new floor to compile for.

CRAFT

"Coding Expertise Is Going to Collapse": The Friction Argument, 505 Comments Deep

Lars Faye's AI Coding Will Prevent Expertise (514 points, 505 comments) is the strongest articulation yet of the "expert novice" paradox: juniors are told to use AI or be left behind, but wielding the tools responsibly demands exactly the expertise the tools let them skip building. Faye's evidence stack is worth reading on its own - the UPenn study of 1,000 students where AI-assisted learners did 17 percent worse than a textbook group while believing they were excelling, the CHI study JetBrains surfaced showing heavy-assistance participants "skipped crucial planning stages" and finished with an "illusion of competence," and the 2026 Anthropic research concluding cognitive effort, "even getting painfully stuck," is likely important for fostering mastery.

The 505-comment thread argues the counterpoint hard: experienced devs are the biggest beneficiaries, every skill pipeline has inefficiency, and some argue the collapse is survivable because verification skill replaces construction skill. Faye's own prescription is the middle path - use models as Socratic sparring partners and interactive documentation, not answer generators, and treat friction as the feature. It is the same trade Armin Ronacher put on the other side of the scale on Sunday: harder languages become cheap to adopt, and the world gets both more slop and more developers. Read together, the two essays bracket the actual design decision for every team: what does the junior track do when the tools are faster than learning, and which tasks are allowed to keep the mentor loop intact.

Why it matters: if you manage or mentor developers, this is the strongest available argument that AI-assisted learning needs deliberate structure - the studies say the difference between a crutch and a tutor is who does the reasoning - which is exactly the tension in our vibe coding guide and in Dan Luu's notes on the real bottleneck.

TOOLS WORTH A LOOK

1. Headlong (OSS, alpha) - Laude and MIT's agent microharness with persistent agency: the agent keeps thinking between interactions, core is under 10K lines of Bash, one-line install, about $1-2 an hour of background thought on GLM or Grok.
2. PicoMQ (OSS) - durable streams over HTTP on plain object storage; a Show HN with no broker and no cluster to babysit.
3. Bookshelf (OSS) - self-hosted eBook library that runs entirely on object storage.

WHAT ELSE IS HAPPENING

• Your executable is a SQLite database (524 points): Farid Zakaria walks through appending a SQLite file to a compiled executable so the binary and its state travel as one artifact - Simon Willison picked it up in his daily column too.
• seL4 confidentiality proofs complete on AArch64 (191 points): Proofcraft's NCSC-funded work finishes the last major security property, completing the formal isolation proof stack for the microkernel on 64-bit ARM.
• LLMs could control their host machines via inference engines (146 points): Boyd Kane argues a malicious model could exploit parser bugs in vLLM or SGLang to escape the GPU host - CVE-2025-9141, where vLLM's tool parser passed arguments to eval(), is the worked example.
• Hot Chips 2026: CUDA targets RISC-V (86 points): Nvidia's requirements for CUDA on RISC-V - RVA23, ACPI, PCIe coherency - with a SiFive server demo planned; Chips and Cheese doubts most existing RISC-V hardware qualifies.
• Apple moves Sign in with Apple relay addresses to private.icloud.com (469 points): new Sign in with Apple addresses ship on private.icloud.com instead of privaterelay.appleid.com later this year, while Hide My Email stays on icloud.com - and Apple tells developers to update email validation and allowlists to accept both domains.
• How Europe is killing makers and micro-entrepreneurs (1388 points, 865 comments): the day's biggest thread by far - Lectronz documents how the CE mark and GPSR compliance paperwork make small hardware businesses in the EU unviable.

---

Every link above goes to a primary source or our sourced coverage. Tomorrow's brief lands when the news does - subscribe to get it by email.