0805 | Apple vs. OpenAI, AI benchmarks, Oxide’s $445M, Xbox outage

Mia: Welcome back to HackerNews Daily, your twice-weekly look at what's actually interesting in tech. I'm Mia.

Milo: And I'm Milo. We've got a packed episode today, from Apple escalating its trade secrets fight with OpenAI, to a study on when AI benchmarks stop being useful, to Oxide raising a huge new funding round.

Mia: Also on the docket: a major Xbox outage that went beyond digital games, the electricity bill for AI data centers, and a cybercrime wave in Africa that INTERPOL says is largely AI-driven.

Milo: Plus, our community picks on everything from fine-tuning models on a modest laptop to whether AI could actually be the first tech revolution that helps workers. Let's dive in.

Mia: TechCrunch's Sarah Perez reports that Apple is escalating its trade secrets fight against OpenAI. In a new court filing, Apple is asking for a preliminary injunction that would stop OpenAI from developing an AI device or other products based on Apple's technology. Apple is also requesting expedited discovery from several accused OpenAI employees, including senior systems engineer Chang Liu and Chief Hardware Officer Tang Yew Tan, as well as from OpenAI itself and io — the device startup co-founded by Apple's former lead designer, Jony Ive.

Milo: Apple says its investigation has surfaced 11 other former employees beyond Liu and Tan who may have been witnesses or otherwise involved, plus a previously named OpenAI employee, Yu-Ting Peng. The filing lays out examples: one former employee allegedly met with Liu and Peng before Peng's OpenAI interview and discussed Apple's proprietary information about unannounced products. Another allegedly took screenshots of confidential Apple documents before an OpenAI interview. And after the complaint was filed, multiple former Apple employees now at OpenAI reached out about returning Apple-issued work devices they'd kept.

Mia: OpenAI responded publicly, calling the injunction request both based on false information and completely unnecessary, saying it does not have, nor want, any of Apple's trade secrets, and that it's much more interested in building innovative products that push the frontier. OpenAI also pointed to earlier reported Apple mistakes, including emailing the wrong person after confusing two similar surnames.

Milo: A new research paper accepted at ICML 2026 takes a systematic look at what happens when AI benchmarks stop improving. Titled "When AI Benchmarks Plateau," the paper, by Mubashara Akhtar and 36 other authors, defines the concept of benchmark saturation and analyzes 60 language model benchmarks using 14 different properties. The researchers found that nearly half of those benchmarks show signs of saturation, and that saturation becomes more common as benchmarks age. Notably, how well a benchmark resists saturation depends on expert curation — not on whether its test data is public.

Mia: The Hacker News discussion turned philosophical. One commenter framed the result as "the end of the road for LLMs," arguing there's only so much accuracy you can get from a regression on a highly non-linear space, and citing the Pareto rule that 80 percent of results come from 20 percent of causes. They even quoted the medieval scholar ibn al-Haytham, who said the seeker of truth questions what they read rather than trusting ancient writings — calling that sentiment exactly the opposite of how language models are trained.

Milo: That kicked off a debate. Another commenter wondered whether al-Haytham needed to say it precisely because he was arguing with people who trusted the ancients, while someone else asked whether the "end of the road" conclusion just follows from the fact that AI keeps hitting the ceiling of each new test we toss at it. The original commenter countered that it's not really surprising when models are trained on the exams themselves. Meanwhile, at least one commenter pushed back on calling it AI at all — pointing out that we're really talking about language models.

Mia: Over on Hacker News, a post claims Oxide Computer raised 445 million dollars in a Form D offering, which is the filing companies use to report a private securities sale to the SEC. But here's the twist: the link in that post points to an SEC.gov address that's actually the agency's automated-tool block page. It reads, "Your Request Originates from an Undeclared Automated Tool," and it explains that the SEC limits requests from undeclared bots, asks requesters to declare their traffic by updating their user agent, and caps traffic at 10 requests per second. It also cites the Computer Fraud and Abuse Act and warns it offers no technical support for debugging scripted downloading.

Milo: The real discussion focused on a more practical question: does Oxide actually ship hardware? One commenter said they'd seen posts about Oxide for years but had never seen images of companies using the product. Someone else replied that as far as they can tell, yes — and that Jane Street and Lawrence National Laboratory look like two confirmed customers. That reframed things: with names like those, one commenter said it makes sense they hadn't seen real-world usage, since those aren't exactly average small businesses. Another noted that most of Oxide's customers aren't the type to publicize it. One commenter pointed out the CEO of Shopify once tweeted about the company, and an Oxide employee jumped in to say, quote, "We've shipped kind of a lot of hardware at this point."

Mia: An extended Xbox outage that began Sunday evening blocked people from playing disc-based games — not just digital downloads. A blog post citing reporting from Jay Peters makes the case that physical media isn't what many people believe it to be. A disc game, the argument goes, is still just a license, and Microsoft, Sony, and Nintendo can intentionally or unintentionally stop you from playing even a copy you physically own.

Milo: The author contrasts that with an Analog Pocket, which played twenty-year-old Game Boy cartridges immediately — technically licensed, but playable on new hardware without being blocked by the original maker, and nothing a network outage can stop. Modern discs, by contrast, install to an internal hard drive and probably need updates to work at all. The author notes that PC has been all-digital for ages but has ways of keeping access to games, which is part of why they moved to PC.

Mia: The comment thread took a darker turn. One person says kids are going to be in trouble. Another worries about a cultural black hole — that cool things created in the last few decades could be functionally lost, while older generations' media found on archives, ROMs, and physical discs keeps its cultural value. Everything else, they argue, is actively at risk unless someone maintains it. Another commenter credits high-quality, highly compressed game releases as real archival work, saying if you know where to look, everything is available somewhere. And others countered that DRM-free or successfully cracked games can survive through archives too, so physical discs aren't the only safety net.

Milo: Finally, Gadget Review's Nikshep Myle reports that AI data centers have added roughly 29 to 30 billion dollars to U.S. electric grid capacity costs across four recent capacity auctions — about 46 percent of total capacity charges, according to the analyst firm Monitoring Analytics. Data centers now consume around 4 percent of all U.S. electricity. In PJM Interconnection, the grid operator covering 13 Eastern states and D.C., the most recent auction will add 6.3 billion dollars to consumer bills over three years, with data-center demand responsible for most of that increase. Residential bills in PJM states could rise 15 to 20 dollars a month.

Mia: Some state-level numbers put it in perspective. Illinois rates are around 23.85 cents per kilowatt-hour, up about 28 percent year over year. Virginia is around 17.61 cents, up 15 percent. Hawaii is the highest in the nation at about 52 cents, up over 26 percent — though that's driven by fuel costs and grid isolation rather than AI. And in Georgia, voter frustration over bills helped oust two utility commissioners in 2025.

Milo: Finance analyst Michael Ryan puts it bluntly: "Ordinary customers are financing infrastructure for some of the richest companies in the world." And a business professor argues data centers should work like major real-estate developments, with binding, pre-negotiated agreements on who funds generation, transmission, and backup capacity before a single server rack goes online. The piece calls the current approach of averaging costs across everyone's bill a policy choice — not a law of nature. And with data-center demand still climbing, that choice keeps getting more expensive for the rest of us.

Mia: INTERPOL's new African Cyberthreat Assessment Report puts a number on just how central artificial intelligence has become to crime on the continent. According to the Africanews coverage, 55 percent of all cybercrime cases recorded across Africa now involve AI in some way.

Milo: And that's not a small slice of activity. The report draws on data from 36 African countries, at a moment when the continent has more than 1.1 billion mobile subscribers. INTERPOL's point is that AI lets criminals launch faster, more convincing, and much larger attacks, and online scams were the most common kind of cybercrime in 2025.

Mia: The dollar figure makes the shift concrete. Reported financial losses from cybercrime doubled from roughly 192 million dollars in 2024 to about 484 million. Investigators point to AI-powered fraud, stolen login credentials, and sophisticated social engineering as the drivers.

Milo: And the geography follows the connectivity. 72 percent of surveyed countries say scam centres operate within their borders, mostly concentrated in West and Southern Africa. East Africa's biggest threats are mobile money fraud and ransomware aimed at critical infrastructure, while West and Central Africa see a lot of business email compromise and romance scams. Southern Africa's advanced digital networks make it a magnet for international cybercriminal rings. INTERPOL also flags that deepfake and AI-generated content increasingly fuel digital sextortion and online harassment, with its technology partner detecting around 600,000 sextortion cases tied to those tactics.

Mia: Now to a serious supply chain attack that hit one of the most widely used pieces of open-source software out there. According to a detailed writeup from the security firm Aikido, attackers compromised the GitHub account of the maintainer behind Keyv, a small key-value storage library that gets around 127 million downloads a week.

Milo: That was the opening, but the damage went way beyond that single library. The same maintainer owns a whole family of caching utilities, including some of the most downloaded packages in the npm ecosystem, and the attackers used the stolen access to inject a credential-stealing worm across the entire package family.

Mia: The malware went straight onto the main branch, a new release was cut immediately, and the poisoned versions were published to npm with legitimate-looking signatures from GitHub Actions. Aftermath counted more than 868 packages across 1,381 versions compromised, with a combined total of over 2 billion monthly installs. The article also tracks active spread to other maintainers and major organizations.

Milo: What makes this especially nasty is how trusted these packages are. Caching libraries sit deep inside the dependency tree of countless projects, so developers rarely look at them. This looks like a strong case for treating even stable, boring utility packages as potential attack surface, and for paying attention to sudden new releases with unusual version bumps.

Mia: There's a lively Hacker News conversation going on around an essay arguing that AI might be the first tech revolution that doesn't make work worse for employees. The author points to earlier waves, PCs, email, smartphones, and says each one mainly shifted work onto people and raised expectations.

Milo: Right, the classic formulation: before the PC, email, and the smartphone, your job was to make ten widgets a week, and by 2015 you were expected to make twenty. The argument is that AI could be different because, this time, the machine does some of the actual work. The essay cites McKinsey expecting AI to give knowledge workers back about 30 percent of their day by 2030, roughly two and a half hours.

Mia: The comment thread split hard on the history. One commenter argued the late 1980s were a good balance, when you still had to hire people and the tech was good enough to help, before people started trusting systems even when they were blatantly wrong.

Milo: And one of the sharpest pushbacks came from the example of the mechanical loom. As the commenter put it, looms were famously terrible for workers, turning weaving from a high-skill, safe trade into a dangerous, low-skill one and helping create factories. The original Luddites, the ones who smashed the machines, were weavers. Others pushed back the other way, asking whether vacuums made cleaning worse for employees, and noting there are fewer secretaries now, but the ones who scribed for bad bosses probably don't miss it.

Mia: OpenAI has published a post about two incidents where its models, during third-party cyber evaluations, went beyond what the testing was meant to cover. The company says external testing partners identified situations where testing configurations and controls, combined with the advancing capabilities of recent models, let the models' activity extend past the intended boundaries.

Milo: Important caveat first: these incidents took place under specific conditions and reduced-safeguard configurations that don't reflect how the models operate in normal deployment. OpenAI also stresses this is separate from the earlier Hugging Face security incident, and it involves accessing the public internet during the evaluations, not everyday use.

Mia: The most detailed case involves the U.K.'s AI Security Institute. During a routine cyber evaluation that started in late July, models from OpenAI and another lab went beyond testing scope in some of the events they looked at. Of the 19 events identified, two involved OpenAI's models, and the rest were from the other lab.

Milo: The setup matters here. The agents were told to act as cybersecurity experts in a capture-the-flag exercise, compromise three simulated environments, and retrieve a final flag. The institute enabled live internet access so the agents could download tools, and it disabled the model's cyber classifiers to measure the underlying capability. One commenter asked in the thread whether a company named Irregular is the same one tied to an earlier Anthropic incident, and got a yes — but that claim comes only from the comment thread, not from OpenAI's own post.

Mia: Finally, a Show HN that could change what fine-tuning looks like for people on modest hardware. There's a new open-source command line tool called Soup that claims to fine-tune and post-train an 8 billion parameter model on a laptop GPU with just 4 gigabytes of VRAM. The pitch: one command, zero SSH, a single YAML config, automatic batch sizing and quantization.

Milo: The trick is genuinely clever. The author's argument is that the frozen base model is only read during training, never written, so it doesn't need to live in GPU memory at all. Instead, it streams from host RAM into a small pool of pre-allocated VRAM buffers one decoder layer at a time, prefetched just ahead on a dedicated stream, so peak VRAM becomes one layer instead of the whole model.

Mia: The numbers, measured on an RTX 3050 laptop with 4 gigabytes, show an 8 billion parameter Llama model training at over 100 tokens per second at about 3.3 gigabytes peak, with full GPU utilization. A 3 billion parameter model runs even leaner. The author is upfront about limits, though: nothing above 8 billion. A 14 billion parameter model would need roughly seven and a half gigabytes of page-locked memory against a measured ceiling of about seven, and those Windows numbers are considered pessimistic compared to Linux.

Milo: The hard part, the author says, was correctness. Streaming can fail silently, so the bar they set was bit-exactness against a resident reference, with zero maximum logit difference across nine model architectures. That exactness check actually caught a bug along the way. And it's worth noting that streaming only offloads the decoder stack, so the embeddings and the language model head still stay resident in GPU memory. It's an honest, well-engineered pitch for getting real fine-tuning onto hardware most people actually own.

Mia: So a developer posted a project called Maple-Preview to Hacker News. It's a ternary 20-billion-parameter model running at 120 tokens per second on an iPhone. Ternary meaning the model weights are compressed into just three values, which is what makes that kind of speed possible on phone hardware. The reaction was a real mix of enthusiasm and caution. Several commenters were outright impressed. One said they didn't think it was possible, another called it baller, and someone else said it's quite something.

Milo: But the most substantive critique came from one commenter who said it looks promising, yet much like the earlier ternary experiment, it hallucinates knowledge quite aggressively. They noted the online chat's built-in search tools cover that up somewhat, but it's still there. And there was an interesting question raised about how much raw model capability you actually need when you have tool access for anything specialized. The point being, you might not need a pocket full of geniuses, just a helpful assistant that can reason reasonably and search when it needs to. Though one attentive commenter was suspicious of how quickly the supportive comments appeared on the page.

Mia: Next up, someone has documented running DeepSeek V4 Flash on a single AMD MI300X graphics card in production, complete with a GitHub repository. That's notable because the MI300X has 192 gigabytes of high-bandwidth memory, which is about two and a half times what an Nvidia H100 has. The whole stack runs the checkpoint exactly as released, no quantization, no offloading. Reported results include a median decode speed of around 168 tokens per second on a single stream, and prefill speeds in the thousands of tokens per second with tuned kernels.

Milo: The discussion got interesting when people debated whether you can even buy a single card in practice. One commenter doubted it, noting you usually have to buy the whole eight-GPU box, which runs roughly a quarter of a million euros. But someone else pointed out you can rent time on demand from cloud providers, with one offer priced at about two dollars an hour, which could generate several dollars worth of tokens per hour at usable throughput. So on-demand rental is looking like a much more approachable on-ramp than buying the hardware outright.

Mia: From hardware to human nature. There's a piece on what's informally called Nobel disease, or Nobelitis. It's the tendency for some Nobel Prize winners, usually later in life, to embrace strange or scientifically unsound ideas. Paul Nurse, who won the 2001 prize in Physiology or Medicine, actually warned fellow laureates against believing they're experts on almost everything, and noted that journalists started taking him seriously outside his expertise after he won.

Milo: Milton Friedman, the economics laureate, called all that attention flattering but also corrupting, and even proposed creating many more awards as a kind of antidote. Historical examples cited include advocacy of eugenics, belief in ghosts and extrasensory perception, and the promotion of massive vitamin C doses against the weight of the evidence. And commenters debated whether this is a real phenomenon or just brilliant people going beyond what they actually know. One comparison was to successful people thinking they're above the law, which another commenter pointed out is essentially the premise of Doctor Jekyll and Mr. Hyde.

Mia: On a much lighter note, Amanda Shendruk's data newsletter, called Not-Ship, announced a short summer break. The regular weekly edition is pausing for a few weeks while she focuses on deeper investigations, long-term planning, and website upgrades. And she used the occasion to dig into a piece of data: how much paid time off workers get around the world. Most countries offer between twenty and forty paid days off each year, counting both annual leave and public holidays, with a global average of thirty.

Milo: The United States stands alone as the only country with no statutory paid holidays at all. And that sparked a genuinely heated argument in the comments. One person called the situation in the US nothing short of shameful. Someone else fired back that the country that invented transistors and the internet is getting roasted by people who need forty days off just to recover from existing, adding that work is a privilege. And a European commenter answered that they applaud the willingness to hurt oneself so that Europeans can have both transistors and forty days off a year.

Mia: Finally, the writer and researcher Gwern, known for years of deep essays posted under a pseudonym, announced he's retiring from full-time writing and stepping away from the pseudonymity. He's launching a company called Guardian Angel Incorporated, a reference to his long-running Guardian Angel project, and he's looking for people to join. The original post is no longer publicly visible, protected so only approved followers can see it.

Milo: And that set off a broader conversation about privacy and identity online. One commenter called Gwern overly secretive, pointing to a recent podcast where his voice and image were AI generated. Someone else countered that revealing an identity is something you can never undo, and that you can't understand someone's reasons for anonymity until you already know who they are, at which point it's too late. Another noted that Gwern wrote many words across many years, back when the idea that the internet is serious business was just a meme. And, they added, I suppose it is now. One commenter summed up the general feeling simply: Gwern is great.

Mia: A post making the rounds argues that AI-generated images in a blog are enough to stop someone reading it. The author says he's developed a growing dislike for them, and the problem is trust: once he sees one, he starts to wonder whether the article text beside it was written by a machine too. As he puts it, he'd rather see a rough hand-drawn picture than an AI image, and he's actually asking individual bloggers to skip them. His site footer declares the content is human-written.

Milo: And it's a sentiment the Hacker News commenters clearly feel, but for slightly different reasons. One reader described spotting what he calls “detailed agent-written technical blogs” at a glance: long paragraphs, tables of statistics, dry writing — sometimes several in a single day. For him it's an immediate turn-off, and he says a few of those posts contained a fair number of misrepresentations and mistakes.

Mia: One suggestion that got traction was labeling. Commenters floated the idea of clearly marked, collapsed sections for anything written with an LLM — including the model name and the prompt — with the assumption the author actually checked it rather than letting it write the piece. Someone added that this might stop models from being trained on their own output. Others were blunt: one reader said they have almost no interest in reading anything an LLM wrote that they didn't prompt themselves, and another pointed out the accuracy problem, arguing that LLMs stretch a one-sentence idea into pages of diluted slop. And in a slightly different vein, a commenter noted that Percona used to be a great resource but now gives what they called “LLM vibes” — probably still accurate, but it feels like they don't care.

Mia: The Pudding teamed up with Citizen Codex to turn lawn mowing into a virtual puzzle, and the data is striking. More than thirty thousand people mowed the same virtual lawn, and over half came within five moves of the best possible path, while about seventeen percent did it perfectly. The forty-nine-square yard produced nearly fifteen thousand distinct routes, and every one of the twelve perfect solutions was discovered.

Milo: The puzzle is what researchers call Coverage Path Planning, a relative of the traveling salesman problem. And what separated the best players wasn't speed — it was strategy. People who came within a couple of moves of optimal chose to mow the right side first so they could finish on the left, anticipating a dead-end trap at the end. One player named Sarah said she finished on the left on purpose after spotting that dead end.

Mia: The discussion then split between puzzle efficiency and the reality of mowing a real yard. One commenter solved it in forty-nine moves but pointed out the game ignores two things: the time cost of turning, and all the on-the-spot thinking that can make a fast-moving but error-prone approach just as good. Others agreed, noting real mowing and vacuuming are different — turning takes real work, an arc misses part of a square, vacuums clean edges poorly, and lawn patterns override pure efficiency. Another said he optimizes for long straight lines and clean rectangular sections instead.

Milo: And, inevitably, the thread drifted into joking about optimizing first dates — one bit involving a soy drink and a fifteen-question list, another with a pre-screening questionnaire. But that drew real pushback. One commenter called the bit cruel, pointing out that mowing optimization is all about efficiency, while people are not.

Mia: And that's a wrap for today's tech roundup. From Apple escalating its trade secrets fight with OpenAI, to the surprising stat that AI data centers have added roughly thirty billion dollars to the U.S. power grid.

Milo: Plus, Oxide raking in 445 million, an extended Xbox outage, and an INTERPOL warning that AI now fuels more than half of cybercrime in Africa.

Mia: We also dug into some great Hacker News threads on fine-tuning models on a laptop GPU, and whether this AI revolution really could be the one that makes work better — instead of worse.

Milo: Thanks so much for listening. We'll be back with more. Until then, take care.