verdverm

Probably the agent workflow traces, including thinking sections, are of main interest. Used in late training for decision making and problem solving strategies.

Sandy Monroe has a business product he sells to Big Auto where he tears down cars, creates a bill of material in incredible detail, and expert notes about the process of how the car is made.

If that's not distillation in the auto industry, I don't know what would be. They all seem fine with, and benefit from, this as an industry.

Aren't consumers benefiting from cheap chinese batteries, EVs, and drones?

Depends on what country you live in I suppose, but likely a spectrum of yes than any outright no. For example, Chinese EVs are using a different battery chemistry and not putting demand pressure on the more expensive chemistry western manufacturers use

Sorry, I don't, still newer to the quantizing side of things since many are already produced by others.

Any of the Kimi 3 titled submissions also have some comments on this theme too

I've got the nvfp4 running on a spark, impressed enough so far I'm going to see if it can replace my qwen3.6 27B usage

Have had issues with the dflash drafter and vllm, crashes, OOMs, broken thinking... so I'm back to the mid teens tok/s

impressive for a small model, but probably only useful if you cannot fit a bigger model (25-35B)

HN is not a jobs board. Please see the HN Guidelines and FAQ linked at the bottom of most pages. There is a special post you can comment on the first business day of each month called "Who's Hiring"

HN is not a jobs board. Please see the HN Guidelines and FAQ linked at the bottom of most pages. There is a special post you can comment on the first business day of each month called "Who's Hiring"

I've found that having good source material (markdown, dependency source, search results for agents) for the models to draw on significantly improves information accuracy. Definitely worth investing in this side of "harness engineering", don't rely on facts burned into weights

initial impressions, great model for coding, probably swapping it out for qwen 27b for a while to long-term test, more sycophantic than any I've run locally myself

HN is not a jobs board. As a YC company that normally follows the job posting rules (including the special ones for YC companies), I recommend you speak internally about the proper channels. The HN Guidelines & FAQ have more info for the non-special cases.

shoe on other foot feelings, re: not honoring international deals

I always thought this should be easier by telling it to write each letter one a new line, then count, any token separator ought to suffice, so much so you'd think they'd have trained in this strategy given tokens make individual letters opaque

basically the same deal as OC Go, lending credence to token commodification

multi turn sessions, they are typically in the high 90% hit rate across all providers without doing much of anything

There are also US based companies like Fireworks serving up the best open weight models with the compliances we need in US enterprise. Depending on the company, they may offer more/different jurisdictions, EU probably needs a Fireworks like company (haven't heard about one, maybe it already exists?)

I hope all models going forward come with a dflash drafter so we don't have to train one up separately.

The tool I've been using, llm-compressor, can quant models that do not fit in memory (use the sequential pipeline)

https://github.com/vllm-project/llm-compressor

my setup to help you on your way: https://github.com/verdverm/quantr

Though it seems these will not be needed as Poolside has published quants & dflash with their models.

I know about it, that's the one I'm referring to. That proposal is insufficient. I was deeply involved in the private data discussions, built a PDS fork / prototype on ReBAC/SpiceDB, and presented it at the first Private Data WG (which no one from Bluesky attended). Bluesky has chosen a path that suits them, not the community at large.

You can learn more about my work here: https://github.com/verdverm/atproto

the community convos here: https://discourse.atprotocol.community/tag/private-data/2

At this point, Bluesky controls the protocol, decided what they want permissioned spaces to look like, and are not entertaining any other proposals (afaict).

I expect there to be an eventual successor that puts permissions at the core of the protocol from the very start.

Fable has been significantly guardrail'd to the point it refuses to work on login pages because they have auth related activities (as reported by another HN user). There are many such cases of people reporting Fable unwilling to work on innocuous tasks

They have introduced https://agenticresourcediscovery.org/ with many other heavy hitters, which looks quite good without deep diving it. I would not bank on this Gemini Enterprise spin because they are try harding lockin (based on the sales calls I had with them and the product documentation)

Also, The New Stack is not a great source, they don't even do their basic due diligence, the very first words from the project README

NOTE: This is not an officially supported Google product.

In what way is Big G betting on this?

I thought they were the ones that are supposed to be making them for me, and to make that digital coinage for me via the ai autonomous business proxy

ATProto's tbd permission system is insufficient for the granularity needed in enterprise. Many of those features needed (like groups) will have to be built in an app view and by proxy be centralized. ACLs are two generations in the past of IAM history

Chat is also not a great modality for the PDS/ATP, Roomy learned this and is building a dedicated protocol and bridge.

yeah, I was surprised when Sundar said they will release it in a month, since training/evals is not so predictable

have we defined what ai limbo is yet?

rumor has it that pro is not so pro in internal evaluations and that there are internal stakeholders holding it up

What about the water breaks for ads? That seemed real Americanization to me

we know that

1. Different people using Ai have different outcomes

2. There is much more to agents than the model

3. Companies are not successful because of the code, look at how many shitty products we endure

Can you explain the basis for your insistence that models matter?

Can you name another technology that determines success/failure rates?

"latest is not a version" ~ Kelsey Hightower

Childhood psychologists are warning about giving Ai to children, very similar to social media and generally screen time.

How does this factor into your business?