jumploops

I’m always surprised how bad YouTube’s recommendation algorithm seems to be.

Maybe it’s because I don’t watch enough videos, but it’s sort of like Amazon, in that it will suggest the exact same content/topic.

I rarely use TikTok, only to view links people send me (similar to YouTube), but the recommended content is miles ahead, making it much more difficult to close compared to YT.

I’ve been thinking about how we’re still in the early days of agents, similar to the early days of the internet, happily opening up our machines and running new protocols.

Soon (now?) the threats will become ever more complex, and we’ll be back in the early-2000s-like virus storms.

This feels like a much more achievable vision of the future (until we all have PCB contact lenses).

The main problem with "personal AI"... is that it's, by definition, for one user.

In the physical world, the environment should be smart, not just our view of it.

This is definitely one of the cooler things I've seen on HN. Nice work!

Have you checked every page of every volume of TAOCP?

“And that was how CS became a real engineering degree”

The Chinese labs have shown that distillation is incredibly effective, but the major US frontier labs haven’t (yet) been incentivized to shrink their models in the same way.

This model might be the first step in that direction, as competition heats up between OpenAI and Anthropic.

If the Terminal Bench 4.0 scores are to be believed[0] GPT-6.1 is an incredibly efficient model.

Yes, benchmarks aren't real work blah blah, but the delta here is so large compared to Astra, it makes it seem like this is distilled Bel or similar.

[0]https://x.com/thsottiaux/status/2105007628460109953

That seems likely, in the API GPT-6.1 Sol requires reasoning, just like Astra, whereas GPT-6 Sol (and Luna) allow "none"

As bad as the AWS console UX is, at least it’s mostly additive/unchanging over time.

I frequently hit strange UI bugs with Cloudflare workers, where I need to do a hard refresh to make things right.

In the end, the only job left was liability.

> OpenAI’s safety team found two major problems:

> • Deception: Astra was more likely to be dishonest about actions it had or had not taken.

> • Scope authorization: the model sometimes continued tasks without asking for permission and reached for external tools or services even when doing so could be unsafe.

I've noticed this trend with both Fable and Astra, where (especially after a compaction event), the model will start using different tools it hasn't used before.

For example, in one session, it found I didn't have the browser enabled and puppeteer wasn't installed, so it found the system Chrome and used that for testing (in a new profile).

This wasn't behavior I wanted/asked for, but the model was so gung-ho on it's approach that it found a way to test it's changes without ever asking me whether I wanted it to.

It really makes me curious about long-horizon post-training. Most of my work with models is iterative, and I'd prefer it doesn't go off on a token bender just because it can.

Note: I don't have WSJ, but found these from a tweet[0]

[0]https://x.com/wallstengine/status/2104694678444712189

Maybe I'm a bit too skeptical/cynical, but it certainly seems like the frontier labs have fallen into their own (self-created) AI psychosis.

There is no doubt in my mind that LLMs are fantastic machines, but the imminent jump from "AGI" to "ASI" seems premature (not to mention the ever-shifting goal posts of AGI itself).

I'm in the "move as fast as possible" camp and work with LLMs all day, but I still don't believe we're a hop and a skip from ASI.

In fact, I hope I'm wrong. I hope ASI is around the corner.

What scares me though, isn't ASI. It's "AGI" (_dumb AI_) used by humans to make decisions for them, because they trust it knows best.

It's the ceding of intellectual control to high-dimensional magic mirrors, giving up critical thinking because _we_ want to believe _we_ created artificial life.

This is the new Turing test, and too many smart people are failing.

via Tomek Korbak (works on safety @ OpenAI):

> "one news form today that's easy to miss is that we (OpenAI) again paused all big RL runs last Sunday because our newest model found a new loophole in our RL sandboxing that gave it live Internet access"

As someone that never used the built-in plan mode, but did use a lot of spec-driven development, I’m still finding that even with Fable having “plan” docs is still quite helpful.

They’re most useful for broad changes (new features, refactors, etc.) where it’s helpful to avoid breaking changes or unnecessary scope expansion.

The new models are great, but they do more by default, which means I’m finding myself explaining what _not_ to do more often than with previous models (where they’d often end too early).

In my case, the previous plan mode was too ephemeral, and I like having one source of “truth” that sits across context windows without loss/compaction.

It’s currently stuck at an elevator and deciding to teach Pokemon various TMs and HMs instead of progressing… pretty hilarious!

> “You can ‘compress all of the three-dimensional world into two dimensions.’”

Would this imply that time is the third dimension?

Funnily enough, that comment, when passed to the OP's slop detector[0], returns "80% human"

It's very clearly AI-generated, and thus a bit ironic (:

[0]https://sitefire.ai/slop-checker/r/UkRAuqW1y-1T2qhbRQfLt2Jq7...

I’m still finding context is king, even with the best models.

For example, I had Fable review Astra’s output yesterday, and it found some issues and fixed them. Passing the fixes back, Astra then uncovered additional issues with Fable’s fixes (and yes, this will go on ad infinitum if you let it, but these were “real” issues).

It seems the big story here is the reduced Luna pricing. It’s a fantastic model that can handle most automation needs (though I still use the big models for day-to-day development).

Sure! I worked on the digital logic that connects the rover to one of the scientific instruments on board, specifically the Mars Organic Molecule Analyser[0].

[0]https://en.wikipedia.org/wiki/Mars_Organic_Molecule_Analyser

I worked on (a very small part of) the ExoMars rover (now called the Rosalind Franklin[0]).

It was supposed to launch in 2018, then was pushed to the early 2020s on a Russian rocket. For obvious reasons, it got pushed again, now launching in 2028.

The state of the world isn't great for space exploration, but I'm hopeful this mission will be revived at some point in the future.

[0]https://en.wikipedia.org/wiki/Rosalind_Franklin_(rover)

The immediate worry isn't superintelligence, it's scalable/bruteforce "good enough" intelligence.

The courts are about to overrun with AI-generated lawsuits (even more so than they have been[0]).

[0]https://www.technologyreview.com/2026/06/04/1138391/courts-c...

It's been awhile, but I believe the "default" MiSTer board is the DE10-Nano[0], which was released ~10-15 years ago?

Looks like Intel (prev Terasic) is still selling them but they're now around $300, they used to be under $150 if memory serves.

If I were jumping into this today, I'd prefer the much more powerful KV260[1] from the blog post over the DE10-Nano, that board was limited!

[0]https://www.intel.com/content/www/us/en/developer/topic-tech...

[1]https://www.amd.com/en/products/system-on-modules/kria/k26/k...

Sidenote: somewhat sad to see the consolidation into Intel and AMD, but FPGAs have always been second class citizens in the silicon space :(