fraywing

Are these models harder to reason about given they produce tokens non-linearly? Especially since we rely on non-latent reasoning to know what the model is "thinking".

So like... do I get access?

This is mostly to be expected.

After Deno did layoffs we saw a consistent decline in announcements and innovation.

Sad day, but unsurprising.

> GPT‑6.1 Sol matches GPT‑6 Astra at roughly one-fifth of the cost

Astra is a pretty impressive model. Excited to try this.

It feels like the only thing left people are building are harnesses? Harnesses of harnesses?

I'm playing around with it for my MCP hiring protocol ojcp[1] and it seems to work very well for signing attestations at the header level.

[1] https://github.com/ojcp-org/ojcp

It might soon be implemented into https://github.com/dakheera47/Job-Ops so yes, indirectly

Yeah, it's becoming challenging. A lot of that is because talent acquisition teams are either too heavily reliant on poor automation or they are the bottleneck in communication because of the application volume. Trying to fix that by making responses more deterministic and a first-class citizen. So if you are using a claw or Claude to apply, you get the ability to poll over the status yourself in a native way.

Ah, yeah. This is definitely more focused on creating structure to help reduce the "ATS black hole" problem as well as solve for candidate quality signals. It's anticipating headless ecosystem for hiring humans.

It's less about "AI said this guy" and more about proof, identity, and hopefully making getting seen as a candidate easier. Right now the signal/noise problem is progressively making that more difficult.

Hey there, totally understand your sentiment.

We have seven other companies involved that are listed on the website and have provided feedback for general use-cases -- given adoption curves are super hard, early traction we have around OJCP is encouraging to me. A lot more to do, but any help or ideas are greatly appreciated.

I'm chatting with other OSS organizations to help make sure this lands in no specific organizational control.

Adding to the topic rather than dog piling onto the website critique:

I think "the context" the author defines also includes harder to unlock peripheral components that makes software products in 2026 defensible.

For example, security is getting harder not easier. Enterprise trust rides on needing to have up-to-date knowledge on data regulation as well.

Relationships, brand maturity, discovery; all of that becomes rarified as we continue to reduce the actual product launch curve -- but also the "harder to do" hygiene stuff that takes you from a "I vibe coded this in two days, good luck" to a "your successful company can count on us" type of hardening has become progressively more of a baseline requirement.

Their server was sleep paralyzed by the HN egregore

I like this and felt like some of it was much more fluid; but was I alone in feeling like the interjected "uh-huh" or "yeah?" moments felt a little jarring?

Almost felt a bit *uncanny valley* for what "natural" conversation is supposed to be like. If the "uh huh" isn't timed correctly, it'll feel like a zoom call with lag.

Brute forcing positive outcomes by spending more tokens until a happy path manifests does not solve the underlying comprehension (and liability) problem.

I fear a world where critical software is stood up with increasingly non-human governed abstraction because it [seems like it] works.

Software engineers as the review terminal in a conveyor of business-led code mass production... coming to a company near you?

It feels like the gap is closing from an intelligence perspective. Or at least doing some kind of log flattening.

Been playing with GLM 5.2 in different contexts. It's less good if you don't max out thinking, but as xhigh it's been able to solve most problems I was throwing at Opus in the about the same amount of time (via OpenRouter).

Wild time to be alive.

wondering how accurate the extrapolated distances are? Like the Rhino and the tree -- I wonder if the approximation could be useful in something like forensics from a security video?

GLM-5.2 has been impressive.

Having my engineers swap over to it from Claude has garnered very little complaint. The lack of multi-modality is a limitation, but using minimax m3 for that isn't super inconvenient.

I'm absolutely astounded that we even have an open weights model that can do 40% of what is shown in here.

I remember making games ten years ago, and it was such a tedious and painful process. This is effectively lightning in a bottle even at a fraction of it's capability.

The next 12 months will be wild (assuming we don't have Chinese models banned by then in the US).

Music and streaming is severely under attack from an effort/commercial viability perspective given tools like Suno[1]

Not exactly saying this is the reason for their sell, but I'd imagine a lot of professional musicians are feeling the desire to exit the industry.

[1] https://www.npr.org/2026/05/02/nx-s1-5804489/music-listeners...

So is this basically a task-specific MoA transformer arch with a DNN that helps make routing decisions? Trying to understand this.

Can we please get a break?

Praying to the security gods.

It seems like we've have non-stop supply chain attacks for months now?

If DoD systems are running on OpenAI infrastructure, you can't just pause them for 6 months during an acquisition. This gets far more complex than just "liquidation of assets".

What happens if OpenAI collapses at this point? Is it just too big to fail given defense contracts and Microsoft?

The Sora sunsetting marked a big shift towards enterprise focus and meeting Anthropic on the enterprise battlefield, but almost all engineers I work with or know are using Claude at this point exclusively.

Anyone seeing differently?

Gave you a star as well. Love the idea and the demo/configurator is very slick.

Looked at the "how it works" -- is there a way to manually specify leaf element sizing? Thinking about dynamically sized elements that don't have markup-level specified sizes yet via bounding boxes.

Very cool, thanks for sharing!

What's the threshold function? Do you have graduating `No --> Partially --> Mostly --> Open`?

Also what's the update cadence?

How do you think the vibe-coding layman audience is using OpenClaw?

I'm observing that there is some kind of status quo bias nearly uniformly being surfaced by the programming community right now.

I myself have feelings like this, as a software engineer by trade.

"We will forever be useful!" As a sounding cry against radical transformation. I hope that's the case, but some of these pieces just seem like copium.

Incredible uptick in supply chain attacks over the last few weeks.

I feel like npm specifically needs to up their game on SA of malicious code embedded in public projects.

>Letting an LLM write for you is like paying somebody to work out for you.

It's worse than this. If someone is working out for you, they still own the outcome of that effort (their physique).

With an LLM people _act_ like the outcome is their own production. The thinking, reasoning, structural capability, modeling, and presentation can all just as easily be framed _as your creation_.

That's why I think we're seeing an inverse relationship between ideation output and coherence (and perhaps unoriginality) and a decline in creative thinking and creativity[0]

[0] https://time.com/7295195/ai-chatgpt-google-learning-school/

As the "agent web" progresses, how will advertisers actually get access to human eyeballs?

Will our agents just be proxies for garbage like injected marketing prompts?

I feel like this is going to be an existential moment for advertising that ultimately will lead to intrusive opportunities like this.