HNHacker News
TopNewBestAskShowJobs

hobofan

8,360 karma · joined March 21, 2014

Jack of all trades (- master of some?)

goisser94@gmail.com

https://hobofan.com

Mastodon: @hobofan@toot.berlin (Twitter: @hobofan)

---

Currently building https://github.com/EratoLab/erato

submissionscomments
hobofan··on You said no MCP
So how do you explain it still being in the "OWASP Top 10" and the "CWE Top 25 Most Dangerous Software Weaknesses"?
hobofan··on Clef: Open-weight decision models, and new RL fine-tuning platform
Closely connected to decision models: A good library to do ranking based on pairwise ranking on multiple attributes. By using a decision model (especially one that can make decisions on multiple fields at the same time) this becomes a lot faster and more powerful. Could make for a pretty nice search reranker as well as prioritizer for many problems.

Of course you can also do ranking one-off with a decision model, but this likely less stable, and by doing pairwise ranking you can also relatively quickly do incremental inserts to the list.

hobofan··on You said no MCP
In practice, the problem with resources is that many resource collections are too large to list exhaustivley, and if that's the case you will need to need to implement a proper search tool anyways (as the Completions utility isn't a good fit), at which point there is little use in also implementing all of that as a resource, rather than `list_`,`search_`,`get_` for a resource.
hobofan··on You said no MCP
That has been one of the most common exploits for decades.
hobofan··on You said no MCP
> Normally if LLMs want to compose multiple operations, they have the perfect tool for this: bash, or whatever other OS shell is available.

I many scenarios, e.g. running the harness server-side, as is the case for chat interfaces, you don't really want to expose OS shell access as that opens up a huge security attack surface.

hobofan··on Pi.dev: You Said No MCP
There is a MCP SEP that outlines multiple ways, that I hope will sooner or later be accepted: https://github.com/modelcontextprotocol/modelcontextprotocol...

While we are waiting on that to become stabilized, we implemented a inspired/co-evolved way to do that in our tool[0], where you mark individual fields in the request/response schema as being file payloads, so that file exchange can be properly orchestrated by the harness and doesn't pollute the context. We just do inline base64 uploads of the required payloads, which in practice we've seen to work quite will until ~100MB files (which is otherwise also the size limit we usually recommend for file processed).

It's annoying that it's not stabilized yet, but for most bigger customers we've seen, they implement 80% of the MCP servers they connect in-house, so doing adjustments to the tool surface, and metadata has been less of a pain for them than we expected.

[0]: https://erato.chat/docs/features/mcp_servers#file-support

hobofan··on 500k facial scans at UK stations yield no arrests, 1 false positive
I would have said that they are mainly doing it to exploit the mechanism as it just serves their profit motif, and everything else is just a byproduct.

However, it appears that when I looked at the list of investors earlier, I've missed that Peter Thiel is among them. So yeah, there are clear political ties.

hobofan··on 500k facial scans at UK stations yield no arrests, 1 false positive
But why is there a need to measure it in the first place and then only react to it?

You already came to the proper conclusion by just thinking about it for a few minutes. Let's just skip building out a surveillance state, and put that money directly into mental health & housing initiatives instead.

hobofan··on 500k facial scans at UK stations yield no arrests, 1 false positive
It's not just politicized propaganda. A lot of unsafety feeling is almost self-inflicted.

In my opinion apps like the Citizen[0] app are some of the worst offenders. They constantly bombard you with scary notifications, creating a constant environment of fear. With that app people are constantly aware of things that are happening in your general area without any real benefit of informing you. It turns the mildly-annoying-but-ignoreable mentally ill person on the street corner from someone that "scares" the 10 people around them into someone that scares 10k people in the area.

Oh, and Citizen is also directly integrated into Axon's Flock-like surveillance product.

[0]: https://citizen.com/

hobofan··on Meta's Muse appears to use an OpenAI model labeled muse-special
Anthropic does the same thing with reasoning signatures. It has already become the standard pattern.
hobofan··on Factorio that you can touch
Not quite Factorio, but Satisfactory has a board game in the works: https://www.reddit.com/r/SatisfactoryGame/comments/1wj4gc6/i...
hobofan··on Tokens too cheap to meter
???

Yes, because it was a largely random undirected process.

hobofan··on Tokens too cheap to meter
> LLMs plateau in ability, in which they will start getting ASIC'd.

They don't need to plateu for that to happen. There are companies already building AI on ASIC, and IIRC they were approach 12 months lead time. A 12 months old frontier model (Sonnet 4.5, GPT-5, Kimi K2) for 1% of the price is still a rather good value proposition.

hobofan··on The current balance of power in open models
As long as you don't have "realtime" workloads, owning the GPUs quickly becomes the economical option. The main cost problems is in e.g. chat applications where the workload is spikey, and users expect an near-instant response, for which you need to scale the GPUs to the highest spikes of the workload.
hobofan··on Warez: The Infrastructure and Aesthetics of Piracy (2021)
NFO files (both image and text) for current releases and ones going way back can still be found on xrel.to
hobofan··on Fujitsu launches made-in-Japan next-generation CPU FUJITSU-MONAKA
Pretty sure it's worded in a way that it applies only to the server, so it should be read as two distinct things:

- Japan-developed 2nm 3D-stacked CPU

- server integrated, developed, and manufactured in Japan

hobofan··on Nvidia announces native GPU programming in Rust
Why? The point of a shader language is to define a function that outputs some graphics. Why should that not be portable between GPUs/CPUs of different vendors?
hobofan··on Claude Cowork and chat are now one Claude
You are not the typical office worker.
hobofan··on A warning about 'model welfare'
I'm pretty sure the message board that was recently swarmed by OpenAI agents to collude on benchmarks would like to disagree.
hobofan··on EU chief opens door for Canada to become 'associate member'
> It's also the third largest economy in the world

If going by GDP (which this claim usually does), the EU as a whole beats out China for second place by ~10%. When going by individual countries Germany alone is the third place behind China, though with a large gap between the two.

hobofan··on The Google Play app review process now regularly takes longer than a week
From what I was able to see from the outside, App Store has had semi-automated reviews for updates in basically forever (+ some random spot checks), presumably based on some introspection that only triggers a human review if a new system API was used.
hobofan··on Introducing System One Models and Jev
That's still ultimately privacy by contract (where you have to trust the inference providers to uphold their end of the deal), rather than privacy by design.
hobofan··on OpenArm: An open-source 7DOF humanoid arm
I think in target-rich environments, the humanoid form is pretty easy to beat.
hobofan··on Charts built for Chat
Not every visualized data point is tied to a binary decision.

e.g. when highlighting a quarterly revenue result, there is a big difference between missing the target by 1% or 50% (and of course the absolute number behind it is also important).

And on a larger scale, trajectory is also something that is easily visualized in graphs, and hard to boil down to a binary marker.

There certainly is a lot of "forced synesthesia", and more looking at charts than necessary out there, especially when decisions are then made more from the gut than based on the data one is claiming to base the decisions off, but a culture of binary-based decision making is just as bad of an idea.

hobofan··on GPT-6-sol appeared on OpenAI API
Why? This just follows the scheme they have established in 5.6, just that GPT 6 introduced Astra as another tier above Sol.
hobofan··on Claude is only available to people over 18 years
I'd wager it's more due to decision making culpability.

Now, for the next "teen committed suicide after talking to AI therapist" headline, Claude can deflect the blame to the parents for not supervising their child enough.

hobofan··on OpenAI Agents API
> Why use a limited version?

Because in most of those API, even many implementations of the Responses API, you lose a lot of control of where your data is going. e.g. an Agent or the Responses API may automatically invoke a tool call that leaks your data to an external service on the internet, without having an option to intervene.

If you want to have control over your data, you have to have control over your harness.

hobofan··on Growing proof that autonomous cars save lives
Photographing bikes or electric scooters does nothing without enforcement through the city. It allows tracking down the person who placed it wrong, but in the absence of fines through the city, the rental companies surely won't start fining their customers and drive them away.
hobofan··on ChatGPT Images 2.5
In a world were there is a significant amount of people that do things just to take photos of themselves doing things rather than experiencing the thing, this may be the lesser evil.
hobofan··on On the Navier–Stokes Millennium Prize Problem
> Improving models in a holistic way sounds a lot like training to me.

I think that's quite a leap. Using de-indetified data to improve the products is what everyone has been doing since the dawn of web analytics.

← PreviousPage 4 of 34Next →