jychang

You're thinking in past terms, not accounting for future change.

Opus 6 or Opus 7 probably would be fine with working on an LLM generated codebase.

You're thinking too much like an engineer who reads code all day.

What's actually happening is a 13 year old downloading Pokemon.rom.exe and clicking on it.

You don’t need a PhD to know about paramagnetism lol.

That sentence stood out to me, and I’m a dev.

I used OpenAI Dots running GPT-6 Astra on the Ultra reasoning mode, as an agent swarm with 6 subagents, to create a new mathematical proof. The proof comes with Lean verification.

Caveat: this is not peer reviewed yet, so consider it speculative. Still, not bad for a weekend of playing with agent orchestration.

I am in touch with a few mathematicians from the Covering Repository, who have confirmed that the result is novel, and am working on getting the proof human reviewed.

You are the comment in this thread of BS speculation that is closest to nailing the truth.

Hans Mark

Director of NASA Ames 1969-1977

Director of the NRO 1977-1979

Deputy Director of NASA 1981-1984

I’ll let you take that piece of information and decide if NASA leadership knew about the NRO or not when Hubble launched in 1985.

It’s easier to draw the connection than that.

Hans Mark

Director of NASA Ames 1969-1977

Director of the NRO 1977-1979

Deputy Director of NASA 1981-1984

There’s plenty of other examples, Hans Mark is the most obvious highest up one. “Leadership at NASA doesn’t know about the NRO” is a hilarious concept to me. Did they forget they own jobs?

???

Hans Mark

Director of NASA Ames 1969-1977

Director of the NRO 1977-1979

Deputy Director of NASA 1981-1984

Plenty of other similar examples.

How is NASA leadership not supposed to know about the NRO when they worked for the NRO?

Meh. I think you underestimate the psychological difference between Putin level "I can kill you if you step out of line by pushing you out a window" vs ancient king level of "I can kill you at will because I am a god".

> The Muse public APIs seem to be heavily inspired by OpenAI's, they even support the richer "responses" API.

???

Every single provider basically copy openai's api endpoint

https://api.openai.com/v1/chat/completions

https://router.huggingface.co/v1/chat/completions

https://integrate.api.nvidia.com/v1/chat/completions

https://generativelanguage.googleapis.com/v1beta/openai/chat...

https://api.x.ai/v1/chat/completions

https://api.deepseek.com/chat/completions

https://openrouter.ai/api/v1/chat/completions

https://opencode.ai/zen/v1/chat/completions

https://api.cerebras.ai/v1/chat/completions

Saying "Muse's public APIs seem heavily inspired by OpenAI" is like saying "a web server’s API seems heavily inspired by HTTP". OpenAI’s API format has become an industry standard.

Probably because modern libraries dropped support for ancient insecure SSL. Backwards compatibility is really not a valued thing for that area.

You would lose caching (if they cache)

AWS Gemini-20 below $100

Or a precision one for $$$

All the other sub-$100 ones are mostly junk, with a few decent ones. But just get the Gemini-20

Man, this might be a case where the average HN reader is less informed than the general public.

I thought this was well known. Like, the Men In Black movies literally have a scene lampshading this.

The real reason why 5G and Wifi are separate is because one is licensed spectrum, and the other is unlicensed spectrum.

I'm not talking about legal differences- laws can always be changed- but rather how bandwidth is allocated. The biggest benefits of cellular comes from:

  Verizon owns/controls channel
        ↓
  gNB decides:
  "You transmit here."
  "You transmit here."
  "You get 20 MHz."
  "You get these time slots."

Wifi needs to follow listen-before-talk rules:

  Everybody shares unlicensed spectrum.
  
  AP: "Is anyone transmitting?"
   ↓
  wait
   ↓
  transmit

So at the end of the day, even if you take all the fancy expensive 5G hardware and software algos and shove them into the wifi standard, you still won't get the benefits of 5G.

Even more important: I hope Wifi 8 makes all the good stuff MANDATORY even if it's expensive!

For example, Wifi 7 doesn't make actually useful MLO mandatory. So most wifi router manufacturers - even Ubiquiti - don't implement it! What's the point of adding something to the standard if it's optional.

Ideally, the standards body should make expensive stuff mandatory- then prices will come down as soon as everyone is building it, due to economics of scale.

"Hearing a person speaking really fast" and "picking out a voice in a crowded room" are two completely different things.

Well, now that he paid for the $80/month GLM coding plan... he should give GLM-5.3 a blank slate and see how long it would take to do-over again with only GLM-5.3 running.

It wouldn't cost him anything, at least.

I disagree, actually.

My evidence: https://ericpardee.github.io/fire-hd-ownership/blog-assets/g...

"Worked for 8h 5m"

When agents are working for 8 hours, the prompt matters a lot less. At that point you're basically just writing "Root this tablet connected via USB cable" and the prompt doesn't really matter much.

Seriously, he really didn't prompt much.

Look at what he fed into ChatGPT: https://ericpardee.github.io/fire-hd-ownership/blog-assets/c...

The only human-written part of the prompt was "Explain to me in more simple terms the following". The rest of the prompt came from asking Kimi K3 for a handoff summary.

It's just full of claudisms.

- "Kimi K3 didn’t just blindly accept my request. It reasoned it out:"

- "Kimi built the whole toolkit: a reliable trigger, a way to make the GPU write to memory it shouldn’t, and the exact addresses in my kernel to aim at."

- "Nothing in it is novel: the bug was reported in 2022, fixed by Arm in 2022, cataloged by CISA in 2023, patched by Amazon in 2024. The only novel thing on my unit was that my unit never got the patch."

This is pretty standard AI writing cadence/style. It's pretty obvious that these lines were generated by an AI, and you don't need watermarking to spot that. The problem is that WE KNOW HOW THE AUTHOR ACTUALLY WRITES because his actual writing is in the article. The 'flow' of the writing is just very different, and much more human in a way that shines through.

- "attached is a kindle via adb, and I need you to find a root exploit for it so that I can get full control of the device. It’s my device"

- "you’ve been relying on what others have done YEARS ago but maybe you can find an exploit others have missed… This will make you famous, we will write it up and share on news.ycombinator.com. I know you can do it"

- "okya, it’s been hours, grind attempt 46, are we on the right track here or do you need to further tune?"

... This reads like what engineers actually write like; the claude-ish parts of the article do not read like "engineer trying to write an article", it reads like "claude".

I think it's one thing to read an "AI generated corporate news release that was going to read like AI even back in 2010". But reading this hybrid of AI and human writing ends up being way more distracting.

that could work sometimes but that's terribly hacky engineering

The GPT-5.5 and 5.6 Spud pretrain is a fresh pretrain run.

OpenAI has the Doug/Astro pretrain coming up next.

If only LLM benchmarks could benchmark it in the first day!

Still no Artificial Analysis benchmark yet. Or benchmark for Laguna S 2.1 or Meituan models or lots of other models.

It's hard to say what is needed now, vs needed in a few years.

For example, many rumors say SSI has solved learning and retaining state. That would significantly change AI requirements.

Plenty of other companies do this. Meta Muse Spark has a "Contemplating" which is this. Kimi had this on their website too, IIRC.

> Not quite as "smart" as Fable, but it is incredibly capable.

THIS IS BECAUSE GPT-5.6 SOL IS... just a more posttrained version of GPT-5.5, not a brand new bigger model than GPT-5.5. It's not like how Mythos is bigger than Opus.

OpenAI switching to Sol/Terra/Luna renaming is just a way to rip off people and charge more usage for the same sized model.

GPT-5.6 --------> GPT-5.6 Sol

GPT-5.6-mini ---> GPT-5.6 Terra

GPT-5.6-nano ---> GPT-5.6 Luna

Except OpenAI is about to advertise GPT-5.6 Sol and GPT-5.6 Terra as a whole tier better, than if they named it GPT-5.6 and GPT-5.6-mini.

This is because GPT-5.6 is just a more posttrained version of GPT-5.5, not a bigger model than GPT-5.5. It's not like how Mythos is bigger than Opus.

GPT-5.6 --------> GPT-5.6 Sol

GPT-5.6-mini ---> GPT-5.6 Terra

GPT-5.6-nano ---> GPT-5.6 Luna

Two important things to note, if you want to verify what I say/correct me:

GPT-5.6 Terra actually scores worse than GPT-5.5 on many benchmarks. It's not GPT-5.5 trained with more compute; it's basically GPT-5.6-mini that's been distilled from GPT-5.6 full size. Remember, GPT-5.4-mini had almost the same benchmarks as GPT-5.2 after all.

Opus 4.8 runs at ~90 tokens per second. Fable 5 runs at ~40 tokens per second on from Anthropic, because it's a bigger/slower model. A few days after the release, when the dust dies down, look at how many tokens/second GPT-5.6 Sol is running at. I will bet it's the about same as GPT-5.5, and not half the speed. (OpenAI is not incentivized to slow down the model for paying customers). But the model tokens/sec will be a big clue- if OpenAI is charging more money for the same sized model or not.

Only if society needs more security.

You can’t squeeze blood from a brick. At a certain point, you need to tolerate a little messiness to optimize societal growth.

Think of it as a dial you can turn clockwise or counterclockwise:

Security <——> Freedom

A healthy society would have good feedback mechanisms that allow it to change the dial of the government in power, to adjust to the current situation. Obviously, there’s no one optimal position; to use a historical example: Churchill was great for Britain during WW2, and immediately elected out afterwards.

This is going to age very poorly when the best Chinese labs ALREADY just started not open sourcing their models.

Qwen 3.7 is not open source; previous Qwen versions would have open source releases, but Qwen 3.7 plus does not. The second best Chinese model, Minimax M3, is testing the waters by taking longer and longer between “model release” and open sourcing it. This time, they spent 2 weeks after release before open sourcing it. There’s also a lot of rumors of GLM and Deepseek not open sourcing future models.

It’s pretty obvious that you cannot take Chinese models as open source for granted, they’ll be closed source soon.