My M5 MacBook Pro, while amazing, is basically relegated to being a thin client with amazing battery life that exclusively SSH's into my desktop PC for work.

I could spawn a Linux VM / use Docker/Podman locally, but dealing with allocating memory, hard drive space and whatever other limitations non-native containerisation incurs isn't worth the squeeze.

Plus I really want to have Vulkan access, both so I can actually play games like a sane person and also use standardized local LLM tooling (MLX is cool but I don't like dedicating time learning things that I can't leverage elsewhere).

If the Mac Mini / Studio range had Linux support, I might actually consider buying one for my home lab. I know others already do and just put up with MacOS - but if I buy a "pro" product, I want it for "professional" usage damn it

Great devices for watching YouTube and video editing I guess?

For example, there is literally no laptop on Earth as nice as the MacBook Pro. Extremely long battery life, insane CPU & GPU performance, unmatched trackpad - but it runs MacOS. Unless you're a professional video editor, the laptop is largely relegated to being a thin client for a Linux machine you SSH into (or spawn a non hardware accelerated VM and demolish performance / power).

How long would it take Apple to develop Linux drivers (even as binary blobs like nvidia)? Probably a couple of weekends. Why don't they? Because walled garden.

The MacBook is lucky in that it's a high profile device so we see projects like Asahi working at it - but again, without good documentation and despite the incredible progress, the Asahi team have spent years and created an alternative for older Apple hardware that still isn't daily drivable (not throwing shade at the project, it's incredible and I am dying to use Linux on my MacBook).

But what about devices with less enthusiastic communities? Pixel 8? Samsung TV? Asus Router? Fitbit Air? D-Link portable 5g modem? My smart blinds?

These devices have no drivers, locked bootloaders, and no community to spend years reverse engineering hardware (that will be obsolete by the time OSS drivers are developed).

I want to see my high end consumer hardware running OSS and until regulators step in to force the "right to repair software" - I don't see that happening and it breaks my heart.

One shot benchmarks that have exhaustive prompts are impressive, but benchmarks that emulate real software development (iterative changes over time with changing requirements) have close-to-zero pass rates.

Because I was using DeepSeek flash, the interviewer was confused at how fast I was getting edits - he'd only used Claude and was used to it thinking for 40 minutes to update a function.

We were building a frontend application and I was able to do that whole guided prompt/review loop. I finished the task with no defects (high quality code) and way ahead of time.

I could tell the questions he asked were prepped for vibe coding with slow frontier models, like "how do you review the code to ensure it is safe to commit" - "make small/reviewable edits and read the code the LLM emits?".

Despite meeting the criteria of a principal engineer in the other rounds, the feedback on the AI round was that "I didn't delegate enough to the AI and was too hands on" lol.

Imagine if someone developed a Linux based mobile OS that could run on it.

Imagine if someone figured out that the USB-C port could do displayport out.

Then imagine plugging that Linux iPhone into a usbc dock or lapdock (laptop with only a screen, keyboard and usbc input) and having a Linux desktop pop up.

The MacBook Neo already proved that iPhone hardware is fast enough for general purpose desktop computing - why aren't we allowed to do that?

Oh right, because the bootloader is locked and (practically) no one outside of Apple can produce hardware drivers without documentation.

It isn't a lack of skill or lack of demand - it's hardware gatekeeping by vendors.

I have over a 15 YoE writing software and guided coding sessions - that is, using an editor like Zed or VSCode with an LLM integrated, writing code how you normally would but using a flash model to prompt away the annoying parts and/or plan - is as productive as vibe coding, produces significantly higher quality, is actually enjoyable, and you actually stay sharp.

Flash models (DeepSeek v4 flash) tend to be so fast that you don't have time for parallel agents, you lock in and rapid fire prompts, building high quality software while incrementally reviewing it as you go. VETO bad edits and try again or rewrite them manually.

By contrast, I have noticed headless agentic coding tends to be an unreviewable black box. The major issues I've found is that, even with a human-in-the-loop, you accumulate defects which compound and eventually you're spending millions of tokens to make trivial changes in a ridgid codebase.

Ultimately, tiny, highly cached, fast models like Qwen's 27b/a3b range or DeepSeek flash are highly capable and relatively inexpensive to run. Hoping people realise we don't need 14 trillion parameter models and I'll be able to buy some ram for my workstation

If I buy an Android phone, I should be able to maintain/update the operating system.

If I buy a MacBook, I should be able to maintain or change the operating system.

If I buy a TV, smart watch, microwave - I should be able to maintain their software.

The only thing prohibiting this is the fact that drivers are obfuscated and/or hardware schematics sufficient to produce drivers are not provided.

IMO, if the EU wants any technical sovereignty, they'd address this as I don't see European companies producing a competitor to the MacBook - I do however see European companies being capable of creating a competing operating system (or Linux) to install on a MacBook, but only if Apple is compelled to allow them.

The same goes for smartphones, cars, and any digital device

That sounds similar to how LLMs have been trained on GPL code and yet output is not (yet) considered derivative work

Does it enable decompilation remasters of classic games?

It feels like AI is a cleanroom laundromat

I joined a company that is an Anthropic shop and I am genuinely shocked.

Sonnet 5 is a little better on long horizon tasks and headless unsupervised agent workflows - but for in-IDE workflows, it's virtually unusable.

I am so used to flipping around my codebase at warp speed with DeepSeek flash. It's so fast and accurate I don't have time for parallel agents. It's a really rewarding workflow.

Moving to Sonnet, you ask is something simple like "split this into a seperate file" "implement this method" "this is my schema, implement a repository for it". It'll spend 30 minutes thinking and charge like $12. And no token caching, what are you even doing Anthropic?

It's unusable.

DeepSeek are in a league of their own

I have my IDE open and a prompt a given task with limited scope - like "write a function that does xyz" "add tests" "create an endpoint that accepts this json" "write a sql repository for this schema" etc.

I use Rust on the backend so if it compiles it works, meaning I don't need to worry about runtime errors - if it looks right, it's right. On the frontend I use Angular, which is about as constrained an environment as you can get in the FE world, so it takes a lot to lead an LLM astray.

So my requirement for high intelligence is pretty low and I am able to review diffs quickly. Deepseek flash is extremely fast so it can render my prompts quickly - while Haiku, Opus and Sonnet are slow and create the same outputs.

I'm guessing the value of Sonnet+ is less hand holding letting you give it larger scopes of work? Giving an AI a high level task and letting it do its thing without needing to review the output?

For around a dollar a day, I am able to more than double my own productivity. What are the frontier models for/what are they trying to solve?

Are they designed to one shot applications? Is the expectation that we want to be able to "yolo" prompt LLMs and have them complete work unsupervised?

Comments URL: https://news.ycombinator.com/item?id=49149599

Points: 5

# Comments: 5

Database is optional, if you add one, you get advanced caching.

It's just a single executable.

    wget https://downloads.getmaxim.ai/bifrost/v1.6.5/linux/amd64/bifrost-http
    chmod +x bifrost-http
    ./bifrost-http

+ credit card charges (+1.5%)

If I was using the routing features it would make more sense - but they charge significantly more for that so the markup is really just for billing consolidation.

The only reason I didn't go for it is I'm not a fan of Python dependency management, Bifrost is just a single executable that uses nearly no memory and is lightning fast.