freshtake

I hear that 100% of code at Anthropic is coded by Claude, so this was caused by Claude. And also, no one but Claude can fix Claude

This

There are different kinds of addiction. The difference is physical vs. mental.

The best example of this is heroin, which has both a severe physical and mental addiction component, and it's the mental addiction that makes relapse so common.

Mental addictions rewire the brain's chemistry, causing the user to seek and only find joy in the substance. This is a better comparison for social media (albeit not as destructive and instantaneously harmful as narcotics)

How old is your child? Younger than 6-8 it's easy to monitor what they're watching and enforce limits. By age 9-10 it isn't just about what they access in the home. Many schools in America are giving kids computer and tablet access, and kids are smart or curious enough to access social media there.

I agree that a big part of this is educating children about these hazards, but that also doesn't mean we should allow these companies to data science the shit out of our attention and will power. Many adults have concerning relationships with social media too -- exposure, pressure, and manipulation are key ingredients that are difficult for anyone to deal with.

This. Also, technology is ever changing, and expecting parents to constantly keep up with feature rollouts on these platforms is unrealistic.

Personal responsibility IS important, but we also don't allow cigarette companies to advertise on billboards with cute characters (remember Joe Camel?)

Short form video is a different beast altogether, and much more concerning. The fact that these platforms don't offer a way to avoid short form altogether is a big issue.

YouTube allows you to "show fewer shorts" but what if you don't want them popping up at all?

AI Slop is the best thing to happen to these platforms - because it will lower trust and engagement as people (hopefully) become tired of inauthenticity. Rage bait is potent when the event in the video _actually_ happened, but when you realize it was AI generated, the manipulation feels even more obvious (though it was always there).

These platforms should also allow users to understand how the algorithm has categorized them, and be able to configure it. YouTube, Instagram, et al. would be safer places for viewers if they allowed users to tell them what they want to be exposed to, and what they don't. Big tech is dodgy about this currently, because the more control the user has the lower the engagement (good for the user, bad for profit).

"I try to understand things as much as I can while planning, even if I’m unfamiliar with the specific technology. If I manage to steer the LLM well, it saves a lot of trouble later on."

This is the most important part of the article. Key takeaway.

Which functions? which projects? If you have 10,000 talented engineers and you're choosing to reduce headcount it's because you don't know how to attack the business and drive revenue in a significant way.

If you're only interested in the gist of the concept and how it can be applied to compression, without the mathematical rigor, here is my go to: https://bertolami.com/index.php?engine=blog&content=posts&de...

This looks cool and could be a much needed step towards fixing the web.

Some questions:

[Tech]

1. How deep does the modification go? If I request a tweek to the YouTube homepage, do I need to re-specify or reload the tweek to have it persist across the entire site (deeply nested pages, iframes, etc.)

2. What is your test and eval setup? How confident are you that the model is performing the requested change without being overly aggressive and eliminating important content?

3. What is your upkeep strategy? How will you ensure that your system continues to WAI after site owners update their content in potentially adversarial ways? In my experience LLMs do a fairly poor job at website understanding when the original author is intentionally trying to mess with the model, or has overly complex CSS and JS.

4. Can I prompt changes that I want to see globally applied across all sites (or a category of sites)? For example, I may want a persistent toolbar for quick actions across all pages -- essentially becoming a generic extension builder.

[Privacy]

5. Where and how are results being cached? For example, if I apply tweeks to a banking website, what content is being scraped and sent to an LLM? When I reload a site, is content being pulled purely from a local cache on my machine?

[Business]

6. Is this (or will it be) open source? IMO a large component of empowering the user against enshittification is open source. As compute commoditizes it will likely be open source that is the best hope for protection against the overlords.

7. What is your revenue model? If your product essentially wrestles control from site owners and reduces their optionality for revenue, your arbitrage is likely to be equal or less than the sum of site owners' loss (a potentially massive amount to be sure). It's unclear to me how you'd capture this value though, if open source.

8. Interested in the cost and latency. If this essentially requires an LLM call for every website I visit, this will start to add up. Also curious if this means that my cost will scale with the efficiency of the sites I visit (i.e. do my costs scale with the size of the site's content).

Very cool.

Cheers

This. Amazon buyer metrics have been tanking for a while. In general if I don't care about the quality I have better and cheaper places to shop. When I know the brand I want, and want predictable quality, I order from the company directly. Price, service, quality, and delivery time are equal or better than Amazon.

A good move, even if many years late. It's a bilateral trustbuster when the same platform that allows commingling and knockoffs then begins tagging legit items as "frequently returned" in the feed.

Can we also get the ability to filter by seller entity country of origin?

Amazon also needs to offer far better tools for buyers to effectively find and attach to brands.

The TL;DR is that the value of renting vs. buying has a lot to do with being realistic and understanding real estate and investments in general.

Buying a house is often an emotionally motivated decision with many important risk factors... Were inspections comprehensive and thorough or did they overlook an issue with the foundation, wiring, plumbing, etc.? Did the buyer understand the required disclosures, and specifically understand what the seller is not obligated to disclose in their jurisdiction? (e.g., in many areas the seller is not obligated to disclose if a child sex offender lives next door). Is the neighborhood up and coming or struggling?

Getting these wrong can easily negate the potential upsides of ownership.

Renting can also be great, but as the article points out, if it mostly just results in more disposable cash, then you may be better off owning (forced savings). Rental properties often cannot be sublet and also cannot be used as collateral or passed on to family members with a step-up in basis. Rental leases also fluctuate with the market, so it's not uncommon in big cities for renters to be paying close or equal prices of their homeowners next door.

Anecdotally I know many folks who have rented their way through, invested wisely, and done well. I also know folks who have moved around the US and always purchased, did their homework, capitalized on tax incentives, and now have a stable of rental properties that helped them become FIRE.

The problem is that our thoughts, opinions, and ultimately actions are the product of our exposure. Social media gives a small number of companies (and their algorithms) unparalleled and unchecked control over our exposure.

We should be educating children at a young age about the benefits and risks of social media. We haven't adapted the way we educate society in light of massive tech changes.

This will likely be a topic that future humans look back on and wonder why we did this to ourselves.

The best engineers are all three, and can turn up or down these tendencies depending on what's required for the project, business, or personal goals. These should not be fixed in proportion over time, as they are each useful in different circumstances.

I spent time at Microsoft as well, and one of the things I noticed was folks who spent time in different disciplines (e.g. dev, test, pgm) seemed to be especially great at tailoring these qualities to their needs. If you're working on optimizing a compiler, you probably need a bit more Einstein and Mort than Elvis. If you're working on a game engine you may need a different combination.

The quantities of each (or whether these are the correct archetypes) is certainly debatable, but understanding that you need all of them in different proportions over time is important, IMHO.

An interesting debate!

A few things to consider:

1. This is one example. How many other attempts did the person try that failed to be useful, accurate, coherent? The author is an OpenAI employee IIUC, so it begs this question. Sora's demos were amazing until you tried it, and realized it took 50 attempts to get a usable clip.

2. The author noted that humans had updated their own research in April 2025 with an improved solution. For cases where we detect signs of superior behavior, we need to start publishing the thought process (reasoning steps, inference cycles, tools used, etc.). Otherwise it's impossible to know whether this used a specialty model, had access to the more recent paper, or in other ways got lucky. Without detailed proof it's becoming harder to separate legitimate findings from marketing posts (not suggesting this specific case was a pure marketing post)

3. Points 1 and 2 would help with reproducibility, which is important for scientific rigor. If we give Claude the same tools and inputs, will it perform just as well? This would help the community understand if GPT-5 is novel, or if the novelty is in how the user is prompting it

I've used all of the popular coding agents, including Jules. The reality to me is that they can and should be used for certain kinds of low severity and low complexity tasks (documentation, writing tests, etc.). They should not be used for the opposite end of the spectrum.

There are many perspectives on coding agents because there are many different types of engineers, with different levels of experience.

In my interactions I've found that junior engineers overestimate or overuse the capabilities of these agents, while more senior engineers are better calibrated.

The biggest challenge I see is what to do in 5 years once a generation of fresh engineers never learned how compilers, operating systems, hardware, memory, etc actually work. Innovation almost always requires deep understanding of the fundamentals, and AI may erode our interest in learning these critical bits of knowledge.

What I see as a hiring manager is senior (perhaps older) engineers commanding higher comp, while junior engineers become increasingly less in demand.

Agents are here to stay, but I'd estimate your best engineering days are still ahead.

I don't think this generalization is quite fair. I'm sure this is true for some folks and their social circles, but for those of us who engineer and know our way around a Home Depot, the capacity is a game changer. I used to have to rent or borrow trucks for my projects.

Not to mention Christmas trees, moving, helping friends out, etc.

If you're focused solely on local network streaming, you can throw most of the features of modern codecs out the window. The trade-off is bandwidth, but if the network can support 100 Mbps, you can get remarkably low latency with relatively little processing.

For example, Microsoft's DXT codec lacks most modern features (no entropy coding, motion comp, deblocking, etc.), but delivers roughly 4x to 8x compression and is hardware decodable (saving on decoding and presentation latency).

Of course, once you've tuned the end to end capture-encode-transmit-decode-display loop to sub 10 ms, you then have to contend with the 30-100 ms of video processing latency introduced by the display :-)

The article captures the two agendas at work. The reality is somewhat dependent on your situation.

If you understand _how_ you should be using AI in engineering, then AI can speed you up because you know what you're trying to build, you understand the fundamentals, and you are the pilot delegating granular tasks. When bugs pop up or requirements change, you'll have the knowledge required to frame and steer the AI to ensure goals are met, and met properly. When AI gets stuck, you'll be able to quickly jump in and work the problem. Your experience will continue to evolve and improve over time because you're plugged into the work and the code. You'll leverage your experience in unexpected ways in future projects, and perhaps your communication skills will improve as well.

Alternatively, if AI is used improperly, it may provide the illusion of velocity up until the point where you realize that your lack of knowledge or involvement actually prevents the AI from making progress. You want to move beyond a simple implementation, bugs pop up that can't be fixed, or new requirements can't be met. You aren't able to dive in yourself, either because you weren't paying attention to the work or were operating too far outside of your expertise. In either case, the progress you thought you were making might actually be a pile of wasted time and technical debt.

The reactive ingredient is passion, not the learning modality. Low motivation is really limiting, no matter how you like to learn. Of course, this topic is hard to discuss in a quantitative way because the number of engineers you come across in your career is a minuscule fraction of the total. Hard to draw broad conclusions like this, but here are a few:

- Formal education is great for foundational concepts (math, hardware, operating systems, compilers, graphics, etc.). Self-taught approaches tend to be goal oriented (I'm learning X because I want to do Y), which can overlook fundamentals that are important. When you don't know what you don't know, having someone to efficiently guide you can save a ton of time, and for some topics, that mentor is a great textbook or teacher.

- Most engineers I know would consider themselves a mixture of formal and informal/self-taught. Again, if you have passion for engineering then you probably like to learn and build, which means you're complementing any formal training with your own tinkering.

I've met and worked closely with amazing engineers and have never found their education style a distinguishing factor. Their passion however, was obvious.

Also, the examples given in the post (Linus, Margaret) were incredibly academic :-)

The issue, unfortunately, is that society has failed to recognize the real dangers of social technologies (social media, social AI, perhaps AI in general). 30 years from now if we're lucky, we'll be watching senate hearings with AI billionaires being asked how they didn't realize their products were so harmful.

This post isn't (really) about AI, it's a political rant masquerading as a tech post with oversimplifications of politics, economic systems, and AI. It's an entertaining enough read, but I'm not sure why it's on HN.

If I allegedly train off of your training, which was trained off of copyrighted content under fair use, we're good right?

Just asking for a friend who's into this sort of thing.

Lol yeah, very familiar

I don't know, I think the point of the example is one of transparent engineering. VHS players could break for any number of reasons but the manufacturers used to put in effort to make them repairable. Obviously a much simpler piece of hardware, but the relative effort felt much greater.

When I used to use Google Wifi, it regularly struggled to connect or establish/maintain connectivity to the outside world, even though my modem was successfully connected. Similar to nest devices, you often have to power cycle them several times to get them into a good state

Until the code breaks and no one can figure out how to fix (or prompt to fix) it :)

This. For folks who regularly write simd/vmx/etc, this is a fairly straightforward PR, and one that uses very common patterns to achieve better parallelism.

It's still cool nonetheless, but not a particularly great test of DeepSeek vs. alternatives.

This exists to an extent today. If you apply your profession to an area that the government deems a societal good (e.g. you work at a public university), it can be discharged after 10 years. This was status quo for decades.

Yep, and I think most reasonable people would agree with this.