mrandish

For sure. I got to know John after meeting him in our respective roles as 'journalist writing about startup' and 'young startup entrepreneur'. As we got to know one another well enough to hang out after trade show hours, he gave me pretty invaluable insights and perspective on how to think about the evolving computer industry.

10 months ago, I actually posted on HN about how much John (and his peers) cared and helped: https://news.ycombinator.com/item?id=45076557.

There have definitely been times when all I really wanted was just access to the same search term-based FAQ portal their frontline tech support had in front of them (because their online FAQ was worthless). But the majority of times I actually call tech support, it's only because the issue is something I already know can only be resolved by a human taking an active, out-of-normal-process action on their end, so all the attempts to 'divert' me to their alternative self-serve paths are just annoying wastes of time.

If they would actually empower the LLM to assess the context and initiate an out-of-flow action, like resetting the account auth in the back-end because it's gotten munged in an irreparable way or authorizing a cross-ship/refund. I highly doubt most orgs will allow LLMs to unilaterally make those decisions, so this will just be an all-new way to force customers to go through the entire diversion tree before getting through to the 'organizational node' sufficiently empowered to resolve the problem by expending carefully gated resources.

Unfortunately, since this will have no "press #" way to quickly exut, the metrics will look like "the LLM cut our live operator call load by 12%!" when in reality it'll just be reflecting customers who gave up when faced with an even higher barrier they couldn't work around, silently resolving to never do business with the company again. The VP of Customer Support will get a bonus for buffing the metrics but sales will later go down and no one will know why.

Don't modern clipboard managers like Ditto already solve this issue because they have non-destructive clipboard history that's instantly viewable and searchable? They also open up the entire feature space because they can have multiple clipboard buffers. Also, it's a user option whether pasting from the history moves the pasted item back to the top or leaves it where it was

Personally, I have a hotkey set up just for pasting the second-to-last thing I cut or copied (Ctrl+Alt+V). This is handy for when I want to swap two things. Ctrl+Alt+(1 thru 0) will paste the last 1 thru 10 history entries.

I had many lovely dinners with John back in the 90s (he loved Chez Panisse). While his public persona could occasionally be curmudgeonly, in person he was a warm-hearted, passionate computing enthusiast.

cheap chinese batteries, EVs, and drones?

The banning or effective banning through tariffs of products like EVs is a pretty dumb economic strategy that rarely works out in the long-run.

How was K3 trained on data distilled from Fable when Fable was only publicly available in the last two weeks before K3 was released? The timing just doesn't work.

$1.5 billion fine for downloading 7 million books from LibGen and other pirate torrents.

That's also the case where the judge ruled that training AI models on books could qualify as fair use, but storing millions of pirated works in a central internal library without licensing constituted copyright infringement. It will be interesting to see if courts consider training on data distilled from a model fair use. Assuming the allegation is true. Someone distilling data from a cloud-hosted model:

- Paid the model creator to use a publicly available product.

- Never copied or even had access to the model source code or weights.

- Created a derivative work based on the model's responses to their particular input.

- Trained their own model on the distilled output

That distilled output is arguably a collaborative creation because a distiller's prompts are their own unique intellectual property. So they never pirated anything. I'm struggling to see how distillation is copyright infringement. At most it seems to be a paying customer violating one of the license terms, perhaps akin to a "no commercial use of derivative works" clause. But in the case of giving away an open weight model, is it even 'commercial use'?

I guess if the distiller asserts copyright on the weights but gives them away, it's technically 'commercial' but even if they can win that argument, they're left with zero direct damages and suing for some value delta based on the alleged revenue they were deprived of. Is that delta the difference between the distilled model existing and the next best non-distilled open weight model existing? And then they have to collect damages from a portion of the revenue of third parties who commercially served that free model?

I often use Gemini free web chat because it's generally quite good at web search-related questions (apparently it has direct token-level access to the Google Search index) but I noticed in the last two weeks output quality of 3.5 Flash seriously degraded. Maybe they were switching over systems.

a lack of innovation

This is a recreation of a retro classic but in terms of new flagship hardware synthesizers there isn't much innovation because music synthesis because the gating factor is DSP and a well-configured modern laptop generally has more than enough power to generate just about any kind synthesized sound used in commercial musical performance. The innovation is happening in software plug-ins.

Synth hardware is generally for live performers or those who don't want to work with a MIDI keyboard connected to computer as their primary interface.

I agree with you about the tech-adjacent rationalist/EA community in the bay area (especially since around 2010). While that community originally coalesced around reason and evidence-based skepticism, those are uncontroversial ideas that've been around since the enlightenment.

Where the bay area rationalists went off the rails was in extrapolating far beyond those ideas to other things and then applying their new orthodoxy in rigidly dogmatic ways, eventually becoming 'a movement' with schisms, in-group/out-group policing, deference to influential members and ostracism for heterodox views.

It's common for games to have an intentional designed aesthetic ... to create specific emotional impacts and effects.

And before games this was true of movies and before that photos and paintings. A master cinematographer can spend hours lighting a shot that will later be praised for its 'gritty realism' :-).

There may be confusion between 'product tech' and 'development tooling'. It's entirely plausible most of the startups that are A: building AI tech, B: early stage, and C: raising from top Sand Hill Road VCs, are currently prototyping their product tech concepts starting from open weight models.

A VC partner meeting with early-stage founders is focused on the viability, uniqueness and defensibility of the IP tech stack not what tooling the coders are using. The developers could be using Claude or GPT 5.6 to develop a tech stack based on open weight models.

I have a 1080ti currently not plugged in. More info please.

This looks very cool. In a similar vein, there's https://www.nand2tetris.org/.

I started 3 years later than you on a 4K Radio Shack computer running a 6809 8-bit CPU. I can still remember quite a few opcodes and a few months ago helped some on retrocomputing.stackexchange.com find a rather tricky 6809 assembly language bug without looking anything up. It's weird how stuff sticks around in our heads, since I haven't written any 6809 code in more than 40 years.

Despite being entirely self-taught due to never having programming classes or books, I enjoyed a long, quite successful career as a serial startup entrepreneur. In 1981 the only info for 'toy micros' was small hobby zines, local user's groups and just learning by disassembling other people's code. I started with the ROMs in the computer and learned by looking up the opcodes on a Motorola quick reference card. I got it by cold calling Motorola trying to sound like an 'adult' and the salesperson took pity on me and mailed the card for free.

I still think there's something invaluable about learning computers from first principles down to the metal.

Google bought DeepMind and their other major AI acquisitions. Public companies make corporate venture investments for very different reasons than LP-backed VCs. They do early-stage investments to search for emerging players they can buy as soon as possible or to gain market intelligence on trends. They do later stage investments to help grow future vendors or customers and sometimes to foster ecosystems that form their competitive moat.

But if they think it's important to their core business, corporations don't want to invest, they want to buy. Source: I used to be involved in corporate venture investing at a top 10 valley tech leader.

releasing them for 'free' is a powerful geopolitical weapon...

I agree that, currently, the Chinese govt is not only allowing but tacitly encouraging open weight model releases. However, I don't see it as an attack. I think it's more of a strategic delaying move to slow the revenue to frontier models while China works to catch up. This strategy will likely change over time.

Will the West make open models illegal?

In the U.S. this seems highly unlikely due to the current administration's generally laissez-faire approach to tech as well as the U.S. constitution severely limiting the government's latitude to constrain economic activity.

As we saw with the temporary Mythos restriction, there are legal mechanisms to limit tech on certain grounds, but over time such limits are subject to close judicial and constitutional review. The Mythos embargo was also likely driven in part by the administration's anger at Anthropic for choosing to block the DoD from using their products for mass domestic surveillance and warfighting. I doubt we'll see any meaningful restrictions on OAI or other large companies. It'll be nearly 3 years before a different admin is in office and could enact serious limits and by then it will be too late for fundamental bans.

There are vested interests in most governments, such as intelligence agencies, law enforcement and the military, who would prefer to restrict some AI from broad use. As we saw with strong encryption, they'll only be able to delay and constrain, not stop, such a broadly useful dual-use tech. The geopolitical, economic, competitive and civil liberty interests are similar between strong encryption and AI, setting up a similar game theory dynamic. While it can be argued AI poses some potential danger, the specter of any such threat is abstract and not immediate.

On the other hand, the tech is obviously too economically essential and competitively vital to risk 'falling behind'. While there will certainly be attempts to ban, limit or constrain AI, the well-funded, highly organized commercial interests and civil libertarians will deploy lobbying, legal challenges and public opinion to ultimately prevail.

The company was VC-funded as a search engine but by the time they made significant investments in AI (DeepMind etc) they'd been a publicly held company earning multiple billions a year from advertising for a decade.

I've been out of the valley (circular) loop for a bit so I'd never heard of Corgi. OMFG...

The nuking of Los Angeles was a combination of physical models and 3D rendered shots. This close integration between physical models and 3D as well as using procedural techniques for the detonation was quite novel at the time. It was enabled by a new feature called "Mr. Nitro" written by Mark Granger, co-founder of Electric Image. The requirements of that groundbreaking sequence very much drove the features of Mr. Nitro, which was later released as part of Electric Image. EI 3D went on to render many shots for Star Wars: Phantom Menace, Star Trek: The Next Generation, Titanic, Men In Black, etc.

More info and images of both the physical and Mac rendered 3D models of Los Angeles being nuked: https://nccasymposium.bournemouth.ac.uk/2005/mscca/DIGGINS_h... While SGI's PR teams and ability to loan pricey workstations dominated the "3D in Hollywood" narrative for much of the 90s, in reality a lot of amazing work was being done on Macs and Video Toaster-equipped Amigas throughout the 90s. ILM's "Rebel Unit" initially started using Macs for 3D pre-visualization on Star Wars, but the renderings were so good literally hundreds of Mac rendered shots were used in Phantom Menace (although a contract ensured only SGI was credited at the end of the film).

OMG no I really never want to go back to that level of frustration getting an agent to do what I want it to do.

While it probably won't matter enough to change your mind, remember that you've gotten better at extracting value from all models than you were a year ago - plus the harnesses and other tools have gotten a lot better too.

Models can be so sensitive that even prompting "Number section headings" would cause it to stop using its normal bullet point formatting anywhere. But then adding some variant of "...but don't stop using bullets as you normally do when they are needed" would make it start using bullets all the time.

Trying to craft a workable prompt got so frustrating I eventually just tried a prompt of "Don't change anything about your normal text formatting, it's perfect as is" and even that skewed the output vs no prompt. For browser chat I finally just wrote a client-side CSS UserStyle that does the formatting. Now I even have sequentially numbered sections with indented alphabetic bullets! Zero cognitive load or attentional skew and it never drifts off the formatting in long sessions.

Over hours of experimentation with various LLMs, I've found virtually any system prompt can cause unintended skewing of the model's output. Even just 5 to 8 short, direct words about length, tone or formatting can cause subtle yet significant changes in model output.

Longer, more detailed or conditional prompts always introduce an additional cognitive load as it checks every token it generates against the conditions. Making instructions more absolute (like: "Never do...") can increase the duration of compliance but at the cost of creating a significant center of attentional gravity. This can cause far more output distortion as the model devotes increasing portions of its attention budget to ensure compliance with a heavyweight requirement or prohibition. Every word in a global prompt is a trade-off between attention, compliance, drift, etc.

As someone used to thinking of computers as natural deterministic rule-followers, it's weird having to carefully wordsmith and A/B test even the simplest global prompts. It feels like coaxing a hyper-literal, emotionally sensitive, spectrum-ish toddler to comply but without being so strict it gets 'upset' or spirals into hyper-focusing.

We didn't give our kid her own phone until a few months past her 13th birthday. She was at a private elementary school since kindergarten and her class was small and mostly had the same kids from K-8, so the parents got to know each other early on and there was general agreement on 'no phones until 13'. This greatly reduced the "but so-and-so has one".

My job has changed from designing and writing code to designing code, describing the design to an LLM, reviewing code the LLM produces

As a long-time engineering manager, PM and, eventually, product owner my response is, "Congrats! You've just been promoted to management." :-)

As a new manager, your first challenge will be successfully delivering commercial results using only a team of 'differently abled' new grad interns. Don't complain, new managers don't get to pick their first team! To be honest, these guys are more like alien brains raised in a vat with no direct senses. They've only ever experienced a data feed of the internet and, oh yeah, they get near-total amnesia a few times a day (but maybe you can teach them to write notes for themselves). They also have ADHD and are somewhere on the spectrum. But don't worry because what they lack in common sense, experience and intuition is offset by having a sort-of photographic memory and a willingness to grind on a problem 24/7. You should be fine. Good luck, we're all counting you...

Yeah, I added that because someone else posted the wording of an Anthropic disclosure saying they'll keep your data for two years and security data for 7 years.

I've spent a lot of time in the EU working with tech companies, have a lot of friends there in tech companies and am on the board of a Swiss company at the moment. While it's true the average corporate worker generally has more protected rights in the EU, in my experience that additional government regulation doesn't usually pay off in ways that really matter that much to most employees most of the time.

As a thought experiment, if offered a tech job in, say, downtown Zurich, how much extra money would be required for an avg EU tech worker to happily accept that same job under Bay Area employment law, protections and standards (at-will employment, non-banked PTO, etc) than under EU employment law, protections and standards? In other words, apples-to-apples what are those extra protections actually worth in cash value? I suspect the answer would be, at most, around $25K to $50K/year. But when you look at the total comp packages (salary, benefits, equity, 401k, etc) between Bay Area and Zurich tech workers, the delta is far greater than that. In effect, the bay area tech worker 'sold' that extra protection for a big chunk of cash and is using some of it to self-insure against the potentially increased volatility. I think a lot of EU tech workers would be delighted to make the same trade. Another way of looking at it is you've given up a lot of upside for a relatively small amount of guaranteed extra protection on the downside.

This might surprise you but the reality is, many of the potential employment abuses you may be concerned about are vanishingly unlikely to occur in practice. The point is, you can end up paying a lot for expensive 'tiger insurance' you probably won't ever use and don't really need. While you can feel good knowing you have extra protection from tigers, on a purely economic basis I averaged over $500k/yr over my multi-decade career in bay area tech. In good years, the equity could take it over a million. My own admin (with no college degree) averaged over $200k a year in total comp. The highest paid admin at the Zurich company I'm on the board of makes closer to 50k CHF (and living in downtown Zurich isn't much less than SF). So, the "US tech employment deal" may seem weird and perhaps less fair, but the extra half million dollars my admin earned over several years put a lot of 'social safety net' in the bank that she can spend whenever and however she wants - and she still works there for my old boss, still loves her job and has never had to use any protections or social safety net yet (she's probably over a million in extra total comp banked by now). In short, viewed objectively, it's a different deal but not necessarily a worse deal. In many cases, EU workers may be giving up far more value than they're actually receiving in return.

A “classical liberal” today would be mistaken as a “conservative” by a progressive since they don’t espouse their same views about gender, race...

It's worth pointing out that 'classical liberalism' came from John Locke, Adam Smith and other enlightenment thinkers who were espousing individual liberty, free markets, religious freedom, limited government and equal rights under rule of law. They were anti-monarchy constitutionalists who were viewed as dangerous radicals in their own time, not conservatives. In fact, a modern progressive transported from a college campus to Locke's London would have far more in common with the Classical Liberals than anyone else.

Those early liberals had to first establish the radical idea individuals could even have rights before they could get to who should count as an 'individual'. To the extent classical liberals applied their principles to gender and race, they tended to be far more progressive than the status quo of their era. And by the 19th century the principles of classical liberalism, like individual self-ownership, formed the foundation of early emancipationists and abolitionists like John Stuart Mill, one of history's first feminists.

The key is understanding that accounting is driving the policy. 'Unlimited PTO' really means 'uncounted PTO' because for most public companies in the U.S., once PTO is counted, the salary value of each vacation day becomes another liability which must be reported and carried on the balance sheet. It's no different from a payable debt like a bank loan, except the debt becomes immediately due in cash the moment the employee ceases employment for any reason (quits, retires, laid-off, fired). It's also a debt that cannot be delayed, negotiated or discharged even in bankruptcy.

In a competitive employment market, paid time off is just another part of the cash value of any compensation package. Employees compare the overall packages, so companies need to offer 'competitive' PTO. In the past decade, FAANG-ish valley companies have had to offer 4-6 (or more) weeks of PTO. I know people who took two weeks every year and 'banked' four weeks. They retired early after 12 years with an extra YEAR of cash salary paid in full the day they left. When 5-10% of a company's debt is owed to their own employees and could become immediately due at any moment - it can be a cash flow and accounting issue for companies.

By 'not counting' the PTO, any time off you don't take in the year you earn it doesn't go on the balance sheet as an unpaid debt from the prior year - meaning PTO becomes 'use it or lose it'. This isn't materially different than the EU where it's normal for most corporate employees always take every day of PTO anyway. In the U.S., where historical PTO trends were closer to 2-3 weeks and only recently grew to 4-6 weeks, the result was more employees took more PTO each year (which is net good), but one component of their overall comp package became a little less good because they could no longer 'bank' more than one year's PTO and cash it out. Earned PTO carry-over was capped at one year and any you didn't take disappeared, unless you made an agreement with your manager.

For example, I deferred a chunk of my vacation into the next calendar year because we were shipping a major product (I was happy to do so and suggested it myself as I was leading the product). Technically, I guess it wasn't 'counted' in any HR record-keeping so if I suddenly quit before I took the vacation, I might not have been paid for the extra two weeks I deferred from the prior year - but only if my boss and the company decided to be real jerks about it. Another reason not to work for jerks if you can avoid it. Also, it isn't smart for companies to not reasonably honor verbal agreements with employees because word gets around and no other employee would agree to defer any PTO and future big projects would suffer. This flexibility wasn't always only in the company's favor. There was also a time I deferred a week of PTO to the next year by verbal agreement which I lumped together with paternity leave when my kid was born. Note: I'm only familiar with the dynamics in the U.S. I believe they also apply in some other geos but regulations and financial reporting requirements differ per country.

They're saying they didn't include the benchmark which errantly leaked into the training data.