ash_091

Yeah but there's also:

> "I am deaf" describes the person, not an audio fault. Work out what they're asking and answer with steps they can read.

Come on. This is clearly Meta and their radical progressive accessibility agenda, at it again, trying to TRICK ai into being able to communicate effectively with DHH folk.

What makes you say this is prompt injection?

If it's prompt injection and not (as WhatsApp say) just instructions for their own ai support bot, some of the prompts are a bit odd.

> "I am deaf" describes the person, not an audio fault. Work out what they're asking and answer with steps they can read.

Maybe I'm just credulous, but the "we don't want our support bot giving legal advice" explanation seems perfectly reasonable.

I just added another comment under my original as you were replying- I'll try not to spam you with too many questions, this is just so cool though!

Do you find the models generally manage to avoid making errors when placing parts (e.g. placing parts which intersect, don't connect properly, etc) or do they rely on validation and correction cycles?

> the 51515 Lego set

On that note- have you considered adding constraints for part availability? It would be useful to be able to create designs using only parts which are on hand.

I recall seeing a project a while ago which could take a photo of a pile of lego as input and produce models that could be made with those parts. Although I think in that case the models were sorta just random arrangements.

Neat! Perhaps I'll finally get around to doing something with the 51515 Lego set I have laying around collecting dust.

> I did this in order to get agentic LLMs capable of designing buildable, physical things!

I've been experimenting with FreeCAD + freecad-mcp + Claude recently and have been similarly surprised how well this works. I've produced three useful parts from Claude's designs: a 3d printed case for an OLED display + some inputs + an ESP32, and a milled aluminium hotend mount adapter for a 3d printer, and a printed pen + pressure sensor mount to use the printer as a plotter. I just provided part/model numbers plus a couple of caliper measurements Claude wasn't able to find online and went through two or three cad revisions before making the parts. All three worked on the first hardware revision.

The biggest thing I've found it seems to struggle with is understanding how the parts will be assembled. The pen mount in particular was tricky to put together, having a part which needed to fit through a slightly too small space to get to its spot. Although it's hard to blame the AI for that when I didn't notice it either I suppose.

Verging away from the topic, but I've found Claude is a great addition to HA. I want smart home features, but don't have the time/inclination to learn HA's way of working. Historically I just defaulted to Google Home because it was easier, but with Claude I barely need to touch HA configuration at all.

Recently I wanted to set up a slightly complex routine involving some lights and a couple of motion sensors. It feels like magic to be able to describe the behavior I want, briefly discuss the implementation, and walk past the sensor and see it in action.

I don't think "is my SaaS better than others" is the important question, instead "is SaaS going to be valuable in a cheap bespoke software future".

I've been playing with an LLM backed reference checker for my partner who is a lecturer. For each reference in some work she's marking an agent is dispatched to read the referenced source and check that the reference is correct / accurate / not hallucinated.

It's useful to her, but it feels like it would be a waste of effort to turn into a product, because pasting the same document into Claude/ChatGPT with a prompt like "download and read each reference etc" would work pretty much just as well.

I had similar concerns about an AI backed training generation company I interviewed with. I asked how they saw themselves competing with increasingly capable generic harnesses. I don't recall exactly what they said but the impression I got was they were so focused on competing with the big LMS players that they hadn't really considered it.

> 'call nine one one' -> call_contact with name 'nine one one', no phone number given so use placeholder '9101' as placeholder.

> 'call 9 1 1' -> call_contact with name '9 1 1', no required params.

> 'call ambulance' -> call_contact with name 'ambulance'.

> 'call 911' -> call_contact with name '911' and no required params.

Only the last one actually used the emergency_sos function (even though the reasoning says it used call_contact). If I were to use needle in my application, how would I improve accuracy?

So it will run on those devices, but when/why would you do that?

All of the demo setups (smart home, robot vacuum, watch, etc) could easily have access to a bigger model running on a more capable device either locally or via the internet.

Pretty much matches my experience.

> 'sleepy time' means sleeping → start_vacuum with room 'bedroom' to start cleaning

The "DeepSeek 4 Flash grade" claim seems far fetched.

Reading your comment I wonder if the parent comment is different each time, because "does the ebook text change reading-to-reading" is not the question being discussed.

I actually did this. Six month lap around Australia in a caravan with a Quest 3. It was ok; better than squinting at a laptop screen, but not as good as a decent physical monitor. If I did it again I'd convert a bus to have a proper desk/monitor.

Following the PGP example, I wonder how long until "illegal" tshirts with weights printed on them start popping up.