chatmasta

Are you aware of extension_control_path? [0] Since PG18 it allows specifying multiple non-$system directories for extension control paths, which makes it easier to adopt composable patterns of extension bundling with immutable images.

CloudNativePG uses this for its Image Volume Extensions feature [1]. A lot of the CNPG team worked on contributing this to PG core because previously the only alternative was baking “God images” at build time (“-full” in this readme) with all extensions in them.

Now with the extension_control_path GUC, it’s possible to “attach” extensions as container volumes at runtime, without rebuilding the image of the container. Maybe you can adopt a similar approach in pglayers.

[0] https://postgresqlco.nf/doc/en/param/extension_control_path/

[1] https://cloudnative-pg.io/docs/1.30/imagevolume_extensions/

If you can’t see the watermark (you called it “invisible”) then what’s the issue?

But it’s available in Claude Code. I’m hoping that’s a typo missing a word or two in the sentence.

Because the interface sucks. It previews in an embedded viewer, exports to pptx, imports into Google Slides, then I can edit it there but if I want to make further edits on other slides, I’m basically stuck because Claude can’t export from Google Slides back to pptx.

Technically cowork writes python scripts to do all the diagram positioning so if I go and make manual edits it’s not easy for Claude to change those back to the script anyway. Easier for me to guide it and it edits the scripts.

The article says that all images have SynthID watermark which is invisible mark indicating the image was AI generated.

Why would you prefer otherwise? Seems like tragedy of the commons to release flood of unmarked AI generation images, among many other risks.

Can it make diagrams that can be imported into .pptx or google slides as actual diagrams (SVG in PowerPoint I guess) i.e. you can move the components, there are real arrows, etc… or does it strictly produce raster images?

Fable is the first model I found that can actually produce such diagrams in slides, in .pptx format that exports cleanly and consistently to Google slides, and is also able to iterate on those diagrams with specific feedback like “center align the arrow with the blue pill on the left which should be vertically aligned inside the dashed database container and make the arrow terminate in the left edge vertical center of the light blue object storage container.”

Mildly surprising they lifted export restrictions for Mythos too. Isn’t that Fable minus the safety layer?

It just needs to work for a few days after bundle release before the mice find out where the cat is hiding. By then it’s too late, the cat already sees the paw prints and droppings into the mouse hole.

Yes, and there is an open source spec [0] that doesn’t require Google/iOS Attestation but “preferably” providers will make their wallet app available on App Stores [1]:

> To ensure that the User can trust the Wallet Solution, Wallet Providers preferably make their certified Wallet Solutions available for installation via the official app store of the relevant operating system (e.g., Android, iOS). This allows the operating system of the device to perform relevant checks regarding the authenticity of the app.

Of course the chances of any important business implementing a side channel option is effectively zero. Maybe some government agencies will offer the option though.

[0] https://github.com/eu-digital-identity-wallet

[1] https://eudi.dev/latest/architecture-and-reference-framework...

Looks nice (slop article aside), but why is VSR Hybrid only benchmarked on Humanity’s Last Exam and not the other two benchmarks (LiveCodeBench and GPQA-Diamond)? Is this an oversight or are the results too terrible to show?

Better question might be will it skip recaptcha?

So you can either wait for every application to do that, or at least make it possible for an LLM to do it… or you can make the LLM use a computer interface that works with every application by definition.

Claude can build UI but it sucks at testing it and iterating on it. Fable showed some improvements in this regard but alas.

I get pretty aggressive about this. I tell them if they expect human effort in review, then they should put some human effort into what they send me. Then I defuse it by ending on a positive tone with some tips about how to prompt better.

It’s worked well so far. Most people only send me that slop once.

As appealing as this is, it will never gain traction without some backwards compatibility with Parquet and wide adoption of query engines to implement that backwards compatible path.

What tool did you use to export the transcript as HTML?

What is inpainting? Everyone in the comments seems to be familiar with the term, and I don’t see it described in the linked page.

If the newer models require more/better hardware then you’ll lose capabilities.

I think you’re better off renting GPU instances and running all the software on those. It’ll be cheaper than Anthropic and OpenRouter but slightly more expensive than electricity and depreciation of hardware.

Just a hunch, but I think the most cost effective “local” deployment method right now is renting GPU clusters by the hour and running all the inference software on them yourself. This will be cheaper than capital expenditure on hardware that will depreciate and become last-gen, and cheaper than OpenRouter pay per token.

The binomial coefficient is derived from Pascal’s triangle, which is a fact I remember but cannot explain, based on some hazy memories of 10th grade math class.

You’re now being forced to optimize your AI cost usage and produce the same results as before you cared about it.

Those face scans are matching to the biometrics you’ve already provided to your own government when you obtained your modern passport.

My passport currently has a broken chip and I’ve been traveling extensively, so I need to go into the border guard queue every time I enter. It’s very annoying. And they take the face scan each time and their computer compares it to past images of my entry.

I verified my age with Apple by clicking one button and Apple said it assumed I was 18 based on the age of my Apple account (2011).

I guess I’m lucky to be in the cohort that avoids the face scans, and I feel a bit dirty about enabling this, but so far — even living in the UK — the privacy concerns have not manifested for me as I thought they might.

To me, the most disingenuous framing of the “protect the children” narrative is not “children can’t access the stuff,” but “adults can access the stuff, once they provide their biometrics.” The default is to deny access.

Mine are project-specific, which is a bit annoying since I’d like it to be global but there are some project-specific additions. Maybe I’ll (ask Claude to) refactor that to be more composable.

It should be a first class feature of the harness, tbh. It kind of is with the /compact [focus] parameter but this is coarse and leaves no record. I find keeping the handoff files in the repo to be useful for historical context and later debugging.

I’ve basically never edited a skill or memory myself. I make the LLM do it as part of the /handoff skill before I clear a session. That also includes pruning existing skills/memories and resolving any drift.

Even the /handoff skill was written by the model…

> theoretically TVs can get an internet connection through HDMI

What?! How on earth would this work?

Also the incentives aren’t exactly aligned. Yes, Anthropic et. al want you to have efficient token usage (because you’ll use it more, and because of some competitive pressure). But it’s not their first priority especially when they make more money with more tokens.

If a tool like rtk improves token efficiency, but has some negative impact on quality, should Anthropic integrate it immediately? Where is the line? This kind of decision is arguably better left to the user.

What they should maybe do, is have a parameter similar to effort level, that allows the user to opt into native features for token minimizing. Make the tools available but leave the choice of the fidelity/savings tradeoff up to the user.

Building @parcel/watcher was always a huge PITA. I also hope they release this package separately because it’d be nice to have an easy to install, on any OS, file watcher that doesn’t require C++ toolchain.

If anyone here is looking to move Greenplum workloads off Broadcom (or unsupported open source), email me miles.richardson@enterprisedb.com — I’m the PM for WarehousePG [0], an open source fork of Greenplum. We’ve got a cracked engineering team working hard to modernize it.

At EDB we’ve forked Greenplum from last OSS into WarehousePG, added over a dozen customers with petabytes of data, and hired a few dozen specialists. We have an extension for Lakehouse connectivity based on DataFusion (with optional offload to Spark including GPU acceleration) to read/write Iceberg. And we have a lot planned for the next version, which you might infer from the name: WarehousePG 19.

[0] https://github.com/warehouse-pg/warehouse-pg

Does anyone have the actual list of 100? Maybe I missed it but I don’t see the link or list in the article.