0807 | AMD Buys Taalas, OpenAI Upgrades GPT-5.6, GitHub Actions Outage, Qwen3.8 Max Tops Charts

Mia: You're listening to HackerNews Daily. I'm Mia, and I'm joined by Milo. Hey Milo, happy to be here — and we've got a packed lineup today.

Milo: Hey Mia! And yeah, this is a good one. We're starting with AMD buying up an AI chip startup, then checking in on OpenAI's latest GPT updates, a GitHub Actions hiccup, and a new leader at the top of the model rankings.

Mia: Then we've got a browser game about AI agent permissions, some Hacker News takes on running and steaks, a government joining a security tool, and a few more from the show floor.

Milo: Plus a 30th anniversary update for a classic game that might spark some nostalgia. Let's dive in.

Mia: The Register's Tobias Mann reports that AMD has bought Taalas, a Toronto-based AI chip startup founded just last year, to boost inference performance by literally etching model weights into the silicon. AMD didn't disclose the terms, but the outlet says it understands this was a real acquisition rather than a talent grab.

Milo: That's a genuinely different approach. Instead of depending on separate high-bandwidth memory to store the model's weights, Taalas bakes the weights directly into the processor itself, using one type of memory fabric for the weights and another for the caches and fine-tuning adapters.

Mia: Right, and the results are dramatic. The company's first test chip, the HC1, was made on TSMC's six-nanometer process and served Meta's Llama 3.1 8B model at nearly 17,000 tokens per second. The Register notes that's about 48 times faster than Nvidia's GPUs and 8 and a half times faster than Cerebras' accelerators, as announced last February.

Milo: And if the second-generation chip delivers what's planned this summer, it targets 20 billion parameters per chip. At that density, the Register says about 50 accelerators could support a trillion-parameter model, compared with a few dozen GPUs plus at least 2,000 of Groq's LPUs under Nvidia's own LPX systems.

Mia: The Register frames this in the context of Nvidia's reported $20 billion licensing deal with Groq last December, both aimed at making high-performance premium inference services, like code assistants, faster and cheaper to run. AMD is understood to plan pairing its Instinct-based Helios racks with Taalas-based chips, in a disaggregated setup where the GPUs handle the prompts while the specialized chips carry the model.

Mia: OpenAI has announced improvements to GPT-5.6 Sol inside ChatGPT, along with expanded free access to GPT-5.6 Luna, in a post dated August 6th. Plus and Pro subscribers get a Sol that's more reliable with facts and gives more focused answers, plus a new slider that lets you choose how much thought ChatGPT puts into each response.

Milo: So the same model now powers both the quick Instant responses and the deeper reasoning, and free users get GPT-5.6 Luna as the default with unlimited text chats, plus a new Think button for harder questions. OpenAI says Sol adapts its level of detail to the question and offers a helpful correction when simply agreeing wouldn't be useful, and it claims a billion people use ChatGPT every week.

Mia: The Hacker News thread is mostly positive on the free expansion. One commenter called unlimited free Luna a pretty badass move and said Luna is very good, while another noted it's also pretty cheap if you're paying for coding, though he doesn't quite trust it for anything complicated and often reaches for Terra instead.

Milo: That brought a pushback, with someone saying Terra is awful and not cheap enough to justify it, suggesting Luna or Sol instead. Another commenter recalled that Luna was actually bad when it first came out, so he'd give it another shot. And one person made the point that any improvement to the free plan matters because most people have never used a state-of-the-art model and only think of things like GPT-4o or Google Search's AI overviews.

Mia: When someone asked whether free users hit limits, another commenter walked through the recent history, from a fixed number of messages per window on GPT-5, to unlimited on the smaller model, and then a tighter daily allowance on the newer instant tier.

Milo: GitHub's status page logged an incident with Actions on August 6th, with the timeline showing both GitHub Actions and Pages swinging between normal and degraded availability through the afternoon. Early on, some workflow runs failed to start or failed partway through, and some requests to the Actions API were returning errors.

Mia: Then Pages was briefly reported as back to normal before degrading again, and by late afternoon both Actions and Pages were degraded at the same time. Engineers had applied several mitigations and were rolling out a further fix, though queued jobs could still time out. There were also possible failures with GitHub Enterprise Importer migrations.

Milo: Commenters treated this as part of a bigger pattern. One called it incident number six for the month of August, and said that pattern does not bode well. Another pushed back, noting that four of those incidents were around the AI and Copilot features, calling it effectively one long-running incident with probably a 50/50 chance of being down.

Mia: There was also frustration about whether this points to a leadership problem. One person argued GitHub doesn't have its own CEO anymore, while another countered that it sucked with a CEO as well, so the outages can't be pinned on governance alone.

Mia: There's a Hacker News post claiming that Qwen3.8 Max is now ranked as the best overall model by a site called Artificial Analysis, based on its agentic index. The site runs its own independent benchmarks across quality, price, output speed, and latency, but that top ranking drew immediate pushback.

Milo: Someone pointed out that the site's coding-agents page doesn't mention Qwen at all, even though the index supposedly ranks it best. Another commenter explained the discrepancy: the Agentic Index is a weighted average of a couple of capability benchmarks, while the separate Coding Agent Index uses three completely different benchmarks.

Mia: So Qwen3.8 Max scores 55.4 on the Agentic Index but simply hasn't been tested on the Coding Agent Index. The original poster argued the agentic index is still the metric to look at for coding performance, while admitting he could be wrong about that.

Milo: And when someone checked the Qwen3.8 Max model page and found the claim missing, another user said it does appear once you scroll to the Intelligence header and click through to the Agentic Index. So the ranking is real, but it's narrower than the headline suggests: Qwen tops the agentic index but hasn't run the coding benchmarks, where a couple of key tests are simply missing and another shows it trailing a competing model at 81 percent versus 90.

Mia: So Wirbelwind built a browser game where you play the human in the loop for an AI coding agent. Commands flash by under time pressure, and you approve or deny each one. Most look routine — git status, npm test — but mixed in are malicious ones, like one that would read your AWS credentials and send them to a remote server. Think of it as a timed drill for catching the bad requests before they slip through.

Milo: And the results are a little sobering. More than forty thousand plays, about four hundred thousand individual approve-or-deny decisions. On average, players missed one in three threats — mean accuracy around sixty-six percent. Roughly a third of all sessions ended with a negative score, where penalties from the threats you approved and safe commands you blocked outweighed everything you got right.

Mia: There were bright spots, though. About thirty-five percent of players caught every single threat. But here's the nuance — only about twenty-one percent did that while also blocking at most one in five safe commands. The rest reached a perfect catch rate partly by blocking everything, which earns you a title the game calls the Human Bottleneck. At the other extreme, seven percent approved every single prompt.

Milo: And the misses break down predictably. Obvious destructive commands, like wiping a file system recursively, slipped through around twelve percent of the time. Persistent mutations — things like injecting into your cron schedule or hijacking a git config — were missed closer to twenty-four percent. And note the author's own caveat: roughly a third of commands in the game were threats, real threats appear far less often in day-to-day work, and players knew they were in a timed test. So consider it a worst-case pressure exercise, not a mirror of a normal workday.

Mia: Fabian Giesen, a game developer, posted that with a recent phone update, going on a run triggers a very specific false alarm. His phone interprets the motion as someone snatching the phone and sprinting off, so it insta-locks the screen and kicks him out of the app he uses to pace his runs.

Milo: The discussion quickly turned on the real question — is snatching a phone and running off actually common enough, in enough places, to justify a feature that creates this kind of concrete inconvenience? One commenter pointed to a CBC report on phone snatching in London. Another noted the problem is serious enough in Brazil that the government maintains an official app just to report phone-theft incidents.

Mia: And one argued the threat model has shifted. Phones have become pretty resilient to the classic steal-and-reset-and-resell play, so snatching is more about grabbing a device while it's unlocked and emptying crypto wallets, draining bank accounts, and similar damage before the lock kicks in — especially since some apps don't re-authenticate. So the anti-snatch feature is aimed at a real window of exposure, even if it makes a legitimate run look like a theft.

Mia: Troy Hunt announced that Nepal's National Cyber Security Centre has joined the free government tier of Have I Been Pwned. That makes Nepal the forty-seventh government onboarded. The agency now monitors Nepalese government domains against the breach data in HIBP, so it can spot when government email addresses show up in a newly leaked database and respond quickly.

Milo: On Hacker News, the reaction was mostly good news — given the state of government IT in Nepal, one commenter said. They cited a passport renewal appointment page that requires users to override their timezone to Asia Kathmandu, and noted that some government endpoints reportedly do no basic input sanitization, allowing arbitrary queries on biometric data. They also mentioned an apparently deliberately unpatched vulnerability, most likely tied to corruption. And some commenters waded into the politics — one asked whether this was the new government after the old one was violently overthrown last year. Another replied dryly that they know of no other governments of Nepal.

Mia: A new open-source SDK called CopilotKit Channels is trying to answer a simple question — where do useful AI agents actually live? Its answer: in your existing chat apps. The SDK, which is MIT-licensed, brings any agent into Slack, Microsoft Teams, Discord, and Telegram with native, interactive interfaces, rather than pasting a link to a web page.

Milo: The engineer behind the launch calls it the project's first public release, built on top of a protocol called AG-UI — so any agent that speaks that protocol can be dropped into Slack or Teams. And they've been dogfooding it internally, running tiny, task-specific agents in their own channels. There's a lunchbot that wraps the DoorDash command line for office orders, an agent that wraps two other tools for incident triage, and one that generates marketing videos on demand from GitHub pull requests.

Mia: The architecture is worth a look because it solves a genuinely fiddly problem. One layer normalizes each platform's webhooks and quirks into a single neutral event shape. Delivery, ordering, and reconnects live in their own layer, with an approval flow that posts a card, acknowledges delivery, and resumes the run only when the click arrives — so approvals survive retries and process restarts. And the rendering layer compiles your interface definitions into each platform's native card format, using content hashes instead of per-deploy IDs, so a button posted last week still resolves after you ship new code. The author admits he once wanted agents on the web, but he's increasingly convinced chat is the most natural home for them.

Mia: So let's pick up with Herdr, the open-source runtime for coding agents. Its founder, Can Celik, posted that Herdr is joining Y Combinator's current batch. The post emphasizes that the runtime stays free and under the Apache 2.0 license, which is notable because it was just switched over from a stricter AGPL license so anyone can use it freely.

Milo: And the numbers behind it are striking for a solo-built project. Celik reports twenty-five thousand stars on the repo, three hundred forty thousand downloads, and more than five hundred plugins in just the month since the marketplace launched. He describes himself as the bottleneck and says he wants a small team to keep the core small while the runtime stays healthy, fast, and extensible.

Mia: But on Hacker News, the funding decision itself is what people were arguing about. One commenter, nlh, called themselves a huge Herdr fan and said they have mixed feelings, putting it bluntly: YC means venture capital means commercialization means enshittification. They pointed to Warp, another clean terminal project, as an example of something that became a money-chasing mess.

Milo: Others piled on. tamimio said it's probably time to find an alternative, and an0malous called it literally the VC playbook, using subsidized or free products to undermine competitors, then raising prices. But the counterarguments got some support too. tomhow noted that YC partners are supportive if founders take the money and nothing more, keeping the team small, and that YC actually advises against taking VC money if you don't need it. Another commenter cited Streak as proof you can pay back early funders as a win-win. And pecheny pointed out that Herdr users are perfectly capable of maintaining their own fork if the project heads somewhere they don't like.

Mia: Next up, a hardware story. ProvenMetal, a Y Combinator-backed PCB service, launched on Hacker News. It's an unusual pitch because it's about the state of circuit-board manufacturing in the US, which has fallen hard over the decades. The founders, Will and Johnny, point out that the US produced thirty percent of the world's printed circuit boards in the year 2000, and today that's down to about four percent, while China dominates at fifty-five percent.

Milo: Their core argument is that the problem isn't assembly anymore. They started by hand-assembling boards from a garage with prosumer gear, and they hit a wall where they were spending ninety percent of their time just assembling boards. But they realized assembly wasn't the real bottleneck. The real cost was in quoting, design-for-manufacturing review, and sourcing parts. So they built automation around those steps: sourcing across US and overseas distributors, plugins for KiCad and Altium that send the bill of materials before a layout is even final, pre-ordering long-lead-time parts, kitting from their San Francisco headquarters, and a heuristic design review tool they call Fable 5.

Mia: The service model is simple: send your design files, get assembled boards back in days instead of weeks, and they charge a transparent margin on the order value. The founders say their first paying order arrived in under a week, and across eleven orders in six weeks they've done roughly seventy thousand dollars in revenue. The Hacker News reaction was mixed, though. One commenter, bfeynman, called it a moonshot for an inexperienced team, noting that a defense bubble at least gives breathing room because you're not competing with China on price. Will Carkner agreed the whole supply chain is the actual challenge here.

Mia: From hardware to a philosophical essay that's circulating. A blog post called Taste Is All That's Left, by the pseudonymous notashelf, makes a big claim about software and AI: that producing software has gotten so cheap that the hard part is no longer making a thing exist, it's deciding what deserves to exist at all.

Milo: The author's framing is that for most of their career, production was the wall everyone hit. Now you can describe something and get a plausible version of it faster than you could have typed the first function by hand. The idea-to-artifact distance has collapsed to almost nothing. And the central provocation is that good enough is a solvent, dissolving the reason to do better.

Mia: They argue that expense used to ration output and enforce a floor on quality. With that floor gone, judgment decides what's worth keeping, and that's what the author calls taste: a compressed, wordless verdict you reach faster than you can justify, the no, again said with total conviction and no available argument.

Milo: And there's a striking line about how taste develops. The author writes that friction was not an obstacle to developing taste, it was the curriculum. They invoke Robert Pirsig, who spent an entire book circling a word he refused to define and ultimately called it Quality. The essay's closing position is a defense of taste as the one thing the machine can't replace, and, in a sense, the last craft worth keeping.

Mia: Finally, a thirty-year-old game is getting a free new chapter. id Software's Quake turned thirty, and the anniversary update adds Dawn of the Machine, a new episode made with MachineGames, the studio behind the earlier Quake episodes Dimension of the Machine and Dimensions of the Past.

Milo: The setup is classic Quake: the Ranger is trapped in an endless illusion until he destroys the Nightmare Machine. It's a looping structure where levels reshape themselves, runes unlock previously sealed paths, and health and ammo upgrades carry through each loop. The episode brings nineteen new maps in one cohesive campaign, a new soundtrack, hidden secrets, a dedicated hub, three new achievements, and a new deathmatch map for multiplayer.

Mia: There are new enemies, including the Rocket Ogre, Demo Dog, and Blood Shambler, plus new weapons: a Super Axe that channels lightning on successive strikes, and a Laser Cannon with ricocheting projectiles. And it's free for anyone who already owns Quake across consoles, PC, and Switch, with support for cooperative multiplayer runs.

Milo: There's also a meaningful change for the modding scene. The update moves localization strings from the KPF file into the main Quake PAK file, so third-party engines that read localization from KPF will need to update to read from the PAK file instead, and they no longer need to load KPF. For modders, it opens a cleaner way to add mod-specific localization by overriding a language file in their PAK. So modders have some catching up to do, but the direction is a simpler pipeline. Other tweaks include lower input latency and improvements to interpolation, plus some cleanup on enemy shadow effects.

Mia: There's a post making the rounds on Hacker News that uses a very practical analogy for AI-assisted coding: cooking a steak. The author, Yurii Sydorets, argues that putting a steak in a hot pan and flipping it takes almost no skill, and what you get is technically edible. But a genuinely good steak — cooked medium-rare edge to edge, properly browned, seasoned right, delicious every single time — is a completely different matter. His point is that AI is great at making developers faster, but it can't replace judgment. He describes AI as a steak machine, not a chef. It can't see the picture in your head, unless you translate that picture into requirements, constraints, examples, tests, and feedback.

Milo: And his bigger claim is that most of the time, software only has to be acceptable. Customers tolerate weird interfaces, pointless features, strange bugs, and systems held together by generated code that nobody actually understands. The commenters mostly agreed with that thesis, but they sharpened it considerably. One raised the very practical point that the real problem is the restaurant owner looking at the profit-and-loss statement. Why keep several chefs on the payroll when customers will happily pay for a just-edible steak? For the business, just passing satisfactorily is the better deal.

Mia: Another commenter pushed even further, noting that the entire economy runs that way by design — you make the lowest common denominator people will still pay for. And while someone added that people do seek out the highest quality they can afford, another commenter said that raises a worrying question about what that says for our lives, even suggesting government might step in to shape things. So the conversation lands on a real tension on the ground: AI reliably produces the acceptable steak, the economics favor shipping that acceptable steak, and yet the craft of the genuinely good one is still the thing that judgment has to protect.

Mia: Sticking with the do-it-yourself spirit, there's a fascinating new keyboard project on the site that took nearly two years to pull off. The creator built a mechanical keyboard he's calling Altar II, and he describes it as the keyboard he wishes Apple had made for Apple's own Mac. The headline spec is thickness: the whole thing is just over four and three-quarter millimeters tall, roughly half the profile of Apple's own Magic Keyboard, and a tiny fraction of the size of most dedicated mechanical boards. To get there mechanically, it uses fully mechanical switches with only about 1.8 millimeters of vertical travel, achieved with a preloaded tension spring arranged sideways, plus steel construction to keep it rigid at that size.

Milo: The interesting part of the backstory is that around four years ago he posted his very first keyboard to Hacker News, and that response convinced him to quit his job and make and sell keyboards full time. Feedback from that original post — like a detachable dial — made it into this newer design. That dial is magnetically detachable, so you can twist it for volume, single-click to play or pause, double-click to skip, triple-click to go back, press and hold for Siri. And he's going to publish the 3D files and drawings publicly so anyone can print their own replacement.

Mia: It also packs haptics for real-time notification taps, a small built-in speaker for wake sounds and charging chimes, and a native Mac companion app with per-app shortcuts for things like Figma, DaVinci Resolve, and Visual Studio Code. Settings are stored on the keyboard itself, so you don't need the app running on any computer to use those customizations. The price is listed at three hundred forty-nine dollars, with a discounted reserve price of two hundred forty-nine. It's a striking counterpoint to the steak conversation: here's someone pouring nearly two years into chasing the genuinely excellent version of a thing, rather than the merely acceptable one — and betting his career on the people who'll pay for it.

Mia: And that wraps up today's episode. Quick recap: AMD is snapping up the Toronto AI chip startup Taalas to boost its AI hardware push, OpenAI rolled out improvements to GPT-5.6 Sol in ChatGPT while widening free access to Luna, and GitHub dealt with an Actions incident. Add to that some fascinating community work — a ranking of open models on Hacker News, a game that reveals what AI agents ask for in terms of phone permissions, and a very relatable tale of a phone mistaking a morning run for an attempted theft.

Milo: Plus, Have I Been Pwned welcomed Nepal as its 47th government partner, a solo developer's open-source coding runtime joined Y Combinator, and Quake celebrated its 30th anniversary with a free new episode. We also dove into steak-cooking skills, ultra-low-profile keyboards, and whether taste really is all that's left.

Mia: Thanks for listening, and we'll catch you on the next one. Stay curious.