Josh

👋 Hi Bluesky! I'm a Staff Software Engineer with a decade of experience building web applications. Love to share and chat about all things programming and webdev.

Had Claude put together a chart of open weight models on the intelligence index that can run on a ~64gb MacBook over the last year. The gap with SOTA is very real but local intelligence good enough to do real work could be on the horizon.

Another day, another model openai.com/index/introd...

Implementation plans should occur only at the highest levels, and those should be so baked into the architecture of the application that nothing needs to be said up front. The agent will simply follow established patterns.

An engineer should be able to review code written by an LLM on their behalf, and that review should count. If your team was satisfied with a single code reviewer before the singularity it should still be so now. Don't create bottlenecks that don't need to exist.

Code changes are now essentially free. You need to rethink your priors with this in mind.

Software engineering is and always has been a process of discovery. You never know less about the thing you're trying to build then you do right now.

Overnight I trained a Jev-like model to output probability distributions over a set of 40 emoji, which runs in RAM at ~150ms per query. This was super easy and cost ~$10. It works pretty well on a weak training set, and with a day or two of work I could make this sing.

It is strange how much LLMs turned out to be able to solve such a wide range of hard problems that would not, instinctively, seem to be problems that a model of human language would be able to solve This is from a Stanford project that let Astra drive a robot in a kitchen tml.stanford.edu/homebody/

This deserves more attention than it got. Realtime ad/clutter blocking that could easily sidestep the Manifest V3 issue that broke ad blockers in Chrome.

mostly agree with this but I think it's more precise to say "we will review high-level artifacts but the vast majority of what is produced will not be that" newsletter.pragmaticengineer.com/p/the-pulse-... x.com/MarcJBrooker...

Command & Conquer: Tiberian Dawn

Opus 5.5 is an incredible daily driver.

I disagree a little bit because I think Astra is worth using, but this is fundamentally correct.

“Most-likely-text generators” that are discovering new math and can one shot web applications.

Grok 4.7 launched and is... pretty bad.

I absolutely love this

Claude Code finally adding support for `AGENTS.md`. Time to simplify some repos…

Watching Jev power computer use/web browsing is wild. It's much faster than current LLMs. If this works I'd use it all the time.

Anthropic models are in a really weird place right now where Sonnet just isn’t cost competitive at any reasoning level with Opus. I can’t figure out when I’m supposed to use this model. Hopefully Anthropic sorts this out soon, because I need something cheaper to run for well defined tasks.

I have somewhat thought here. I wonder if this is something that your orchestrating agent would invoke rather than you directly. 

Jev: Fable-level model that doesn’t charge for output tokens because they’re too cheap to meter it’s not general though, it only makes decisions, doesn’t generate text, but input tokens are measured by the billion ($42/btok) typesafe.ai/blog/introdu...

$0.042/million tokens in and zero cost output tokens is insane. The speed is wild too. Can’t wait to try this out.

typing `/effort xhigh` into slack DMs to get better answers from colleagues

Good luck building out multiple redundancy, offsite backups.

Microsoft peaked with Windows 98 SE.

Every time I log into Minecraft I'm reminded how much I hate Microsoft. How you can mess up auth this bad in a application nearly two decades old is truly an accomplishment.

Ultimately it's not the code we care about, it's the product.

Does a non-apple laptop exist that's competitive with MacBooks in battery life, build quality, and noise? Between that and unified memory I haven't found any good options.

In 2027, mysterious hacks attributed to open source weights and definitely not connected to the companies that accidentally hacked a bunch of companies in 2026, which was okay

Hot take but most of the human-authored production code I've encountered in my career was significantly worse than what any of the frontier models produce today, and we thought that code was perfectly fine.

"If we execute these measures well, I believe they would slow China’s progress enough to widen America’s lead significantly over the next 3–5 years — the window when AI becomes geopolitically most important." So 3-5 years is the real horizon for AGI/ASI.

Looks like Anthropic, OpenAI, and xAI have all agreed to "pace" frontier models. The steps proposed by Amodei look like a nuclear arms limitations treaty, complete with third party inspectors.

So, Moonshot was pretending to serve Kimi but really serving Claude? Wild stuff here.

Mostly good takes although I disagree with the idea that the death of middle management means the death of career development. One of the biggest mistakes our industry made was making management the default upward path for software engineers. 

Independent of policy debates about existential risk, it is worth noting that even if AI development stopped today with the models we have right now, we would still have at least a decade of roiling change throughout much of work, education, and social life as harnesses improve & AI diffuses further

I've been rethinking the trade offs of typed/compiled languages vs. dynamic/interpreted ones. LLMs have significantly reduced the costs while offering additional benefits as software grows in complexity faster.

We need a Manhattan Project for AI alignment.

My current orchestration flow looks roughly like this. Anyone doing anything different?

Anthropic's take on what AI is going to do to jobs and the economy over the next few years. Apart from the (mostly) reasonable projections, there's some good visualizations of how roles are being transformed.

I had an idea for a project I wanted to do and Astra actually talked me out of it. Using it feels like it knows what I'm trying to get at and can help me down the right path, even if it's a different path than what I initially envisioned.

Had Astra reskin Spotify to look like Winamp cause why not