AMD Acquires Taalas — Etching AI Models Directly Into Silicon
857 points · The Register
AMD has acquired Taalas, a Toronto-based startup that takes the concept of “hardware-accelerated inference” to its logical extreme: instead of storing model weights in HBM memory, Taalas etches them directly into the silicon itself. Their first test chip, the HC1, fabbed on TSMC 6nm, served Meta’s Llama 3.1 8B at 16,960 tokens per second — roughly 48x faster than Nvidia GPUs at the time of its February reveal.
The approach creates what Taalas calls “model-specific integrated circuits” (MSICs). Each chip has a mask-ROM recall fabric where weights are permanently baked in, plus SRAM for KV caches and fine-tuning adapters. The second-gen HC2 aims for 20B parameters per chip, meaning you’d need just 50 accelerators for a trillion-parameter model. AMD’s plan appears to be pairing these with their Instinct GPU racks in a disaggregated architecture — GPUs handle prompt processing, Taalas chips handle token generation.
The catch is obvious: you’d better be absolutely certain about your model before you fab it. Weights in ROM mean no post-training updates without a new tape-out. This is the semiconductor equivalent of burning a CD-R. For stable, high-volume inference workloads (code assistants, API endpoints serving a single model), the economics could be compelling. For anything that needs to iterate, it’s a very expensive commitment. AMD didn’t disclose the price, but the strategic logic is sound — they need asymmetric advantages against Nvidia’s CUDA moat, and “17K tokens/second per chip” is a hell of a talking point.
New Mexico Court Orders Meta to Pay $942M Total Over Children’s Mental Health
676 points · The Guardian
A New Mexico court has added $567M to the $375M fine Meta already owes from its March trial loss, bringing the total to $942M. The original jury verdict found that Meta knowingly harmed children’s mental health and concealed what it knew about child sexual exploitation on its platforms. This latest ruling — the second phase of the landmark trial — specifies how the money gets allocated: $420M for treatment services for young people, with the rest going to awareness, prevention, and screening over five years.
The judge also ordered operational changes: Facebook and Instagram must build banner screens explaining protection features and tools for handling inappropriate content. Meta must develop an “under-13 prediction model” within two years and request proof of age for users it estimates to be minors. Interestingly, the court declined to impose blanket age verification, noting that doing so for Meta alone — while competitors face no such requirement — would be “inequitable and unduly injurious.”
This is the first trial to find Meta liable for acts committed on its platform, following a 2023 Guardian investigation that exposed Facebook and Instagram as marketplaces for child sex trafficking. Former moderators testified they flagged harmful content that was never escalated. Nearly a billion dollars is a real number, but Meta’s annual revenue is north of $130B. The question is whether the operational mandates — the prediction model, the content screens, the state-reviewed educational campaign — will actually change behavior, or whether they’ll become another compliance checkbox that Meta’s lawyers navigate around.
Taste Is All That’s Left
640 points · notashelf.dev
This essay argues that AI coding tools have collapsed the “idea-to-artifact distance” to near zero, and in doing so, have stripped away everything that was never the actual skill. The typing, the syntax, the wiring — all of it was always, in principle, automatable. What remains is taste: the compressed, wordless judgment that tells you two of three plausible implementations are wrong before you can articulate why.
The mechanism is uncomfortable. Taste is built through friction — shipping bad code, living with its failures, filing the failure away. Remove the friction (by offering competent output for free on day one), and you get developers who are more productive by every measurable metric but who never developed the judgment that the old process forced. The essay invokes Pirsig’s Quality, Sturgeon’s Law, and the Arts & Crafts movement to make its case: when making becomes free, choosing becomes the craft. Curation was a minor virtue when things were expensive to make; it’s the whole game when they’re free.
The economics are cruel — taste is slow, invisible on a dashboard, and you ship at exactly the same speed as someone without it. The essay also includes a candid post-mortem acknowledging the HN front page hit and that commenters flagged it as “AI slop.” The irony of an essay about taste being accused of lacking it is either deeply meta or a genuine problem. Either way, the core argument stands: effort was a filter, we removed it, and now judgment is the only scarce thing left.
A Year of Fighting Scrapers on My 1.5 Million-Page Website
311 points · patronview.com
(Cloudflare blocked direct extraction; summary based on title, HN discussion, and context.)
A website operator reports that 99% of their traffic is now bots — primarily scrapers feeding AI training data and search indexers. The site, which hosts a curated collection of information, has been dealing with this for a year. The HN discussion highlights the core tension: the site itself is built on scraped and curated data, making it a particularly attractive target for re-scraping. One commenter noted that “AI services currently treat the entire web as their storage and cache layer.”
The 99% figure is striking but not surprising to anyone running a content site in 2026. The bot economy has exploded — AI companies, SEO tools, price scrapers, and content aggregators all hit the same endpoints with different user agents. The practical costs are real: bandwidth, compute, database load, and the constant arms race of blocking and re-routing. What’s interesting about this story reaching the front page is that it articulates something site operators have been whispering about for years: the open web is being eaten by its own parasitic traffic, and the economics of running a content site are getting worse, not better.
São Paulo Resident Transforms Degraded Area Into Urban Forest
301 points · São Paulo Secreto
In the early 2000s, Hélio Silva — a resident of São Paulo’s East Zone — started planting trees on an abandoned, garbage-strewn strip of land along the Tiquatira stream. He spent his own money on 200 Atlantic Forest seedlings. A local businessman, who used the space as a parking lot, destroyed all of them. Silva bought 400 more. Those were destroyed too. He bought more again, and eventually the community rallied behind him.
In 2007, São Paulo’s City Hall formally established the area as the Tiquatira Linear Park — the city’s first. Today it covers 320,000 square meters with 32,000 trees of 160 species, including jacarandas, jequitibás, and jatobás. It has sports courts, walking tracks, skateboarding areas, playgrounds, and kiosks. All because one person refused to stop planting.
Scientists Discover Kelvin-Helmholtz Instability on the Sun’s Surface
298 points · National Solar Observatory
An international team using the NSF Daniel K. Inouye Solar Telescope — the world’s most powerful solar telescope, perched near the summit of Maui’s Haleakalā — has discovered Kelvin-Helmholtz Instability (KHI) in the Sun’s photosphere. KHI occurs when two fluids slide past each other at different velocities, creating small swirling vortex patterns that look like breaking ocean waves. Published in Nature, this is the first experimental confirmation of a phenomenon long predicted by theory but never directly observed on the Sun.
The discovery may help explain why the Sun’s outer atmosphere is so much hotter than its surface — a long-standing puzzle in solar physics — and how magnetic energy builds up and redistributes to fuel solar flares and coronal mass ejections. Those are the events that send energy bursts toward Earth, disrupting satellites, power grids, and communications. The finding also validates the Inouye telescope’s capabilities; no previous instrument had the spatial resolution to see these tiny swirls at the boundaries of magnetic regions.
Oracle Bans AI-Generated Code from OpenJDK
233 points · Dealroom.co
Oracle has banned AI-generated code from OpenJDK contributions, citing safety, security, and intellectual property risks. Developers can use LLMs privately for debugging and reviewing, but cannot submit AI-generated material to repositories, pull requests, or other project channels.
The hypocrisy is rich enough to farm. Larry Ellison has publicly declared that AI models now write Oracle’s own code. Co-CEO Mike Sicilia credited AI tools with enabling smaller engineering teams to deliver faster. Oracle is spending $70B this year on datacentre expansion. Yet for their open-source Java project, AI code is apparently too dangerous to accept. The likely explanation isn’t principled caution — it’s IP risk management. Oracle doesn’t want to end up in a lawsuit over GPL or Apache-licensed code that an LLM regurgitated from a training set. For their proprietary products, they control the legal exposure. For community contributions to OpenJDK, they don’t.
S&P recently downgraded Oracle’s credit rating to BBB-, one notch above junk, citing uncertain returns on their AI infrastructure spending. The ban on AI-generated OpenJDK code is a footnote compared to that $70B bet, but it’s a revealing one: Oracle wants to eat the AI cake while managing the legal risk of who baked it.
The Throughline
Today’s front page tells a single story about the tension between AI capabilities and their real-world costs. AMD is etching models into silicon to make inference faster and cheaper. Oracle is simultaneously betting $70B on AI infrastructure while banning AI code from its open-source projects. Meta is being held financially accountable — to the tune of nearly a billion dollars — for the harms its platform enables. A website operator is watching 99% of their traffic get consumed by scrapers feeding the same AI ecosystem.
And underneath it all, the “Taste Is All That’s Left” essay articulates what all of this is converging toward: when generation becomes trivially cheap, the only scarce resource is the judgment to know what should exist. The São Paulo forest story is the counterpoint — one person with taste and persistence, planting trees that outlasted every obstacle. The solar physics discovery reminds us that some things still require instruments we’ve never built before, pointed at problems we’ve never been able to see.
The AI economy is not just about model performance benchmarks. It’s about who bears the costs — the website operators paying for bot traffic, the children harmed by algorithmic engagement, the open-source projects managing IP risk, the communities living next to $70B datacentre expansions. The front page is asking: when the dust settles, who’s actually better off?