tryingEveryThing retweeted
Big gap!
Pointing on a computer UI clearly has been more trained in the big models.
Specialized models mostly fail at this because they don't have the vocabulary (yet), returning no points.
Pointing on a computer UI clearly has been more trained in the big models.
Specialized models mostly fail at this because they don't have the vocabulary (yet), returning no points.
1
3
18
3,893
Many thanks to @VictorTaelin and @badlogicgames a dream for the last three years is coming nearer. Thanks to orgMem, pi-durable, bend2 and my work over the last three years... This was created in just 1 day. And it is already working full steam ahead!
2
0
4
398
Also thanks to @Prof.Binner for MITO and @Socracity3.0 as well @unclebobmartin for clean code and @GiancarloGuizzardo for UFO ontology.
0
0
0
60
uh, THIS is just perfect timing. trying it out right now.
`pip install cactus-needle`, `needle.transcribe("clip.wav")`.
A 16kHz WAV or raw samples need nothing else.
`needle whistle playground` transcribes from the microphone in the terminal.
`needle whistle compare` runs the same clip through Whistle, Whisper and Moonshine with their timings, so you test and use either model.
Code at github.com/cactus-compute…
0
0
1
45
tryingEveryThing retweeted
kv cache is a mean master.
please, someone make models use a context-chunk architecture where chunks are order-independent, so renewing one chunk doesn't invalidate the cache for all the others. maybe some kind of pre-embedding of each context chunk, plus a sort of late interaction between the decoder and those chunks?
please, someone make models use a context-chunk architecture where chunks are order-independent, so renewing one chunk doesn't invalidate the cache for all the others. maybe some kind of pre-embedding of each context chunk, plus a sort of late interaction between the decoder and those chunks?
6
2
26
1,856
who is building background actions in @ArchUpdates + niri? or is that something that is already possible but I am not aware of it?
0
0
0
43
has anyone came up with using jev for analysing usage data of app to efficiently and effectively push new features / update an progressive UI to expand the horizon of the user?
0
1
1
71
just stumbled upon this. wow!
Thanks for this: github.com/programasweigh…
Thanks for this: github.com/programasweigh…
You can describe an AI function in English, compile it once, and run it with Wi-Fi off.
That’s what we’ve been building with ProgramAsWeights: small neural programs you can call like Python functions, locally on your own CPU.
Code and models are public:
programasweights.com
0
0
2
78
have you tried to include a different backend @MaximeRivest into aiconvo? like ax agnt with his graphjin?
ax agent is a great foundation for domain-specific agents because you can specialize the entire execution loop:
• DSPy-style signatures define your task as typed inputs → structured outputs, with enums, nested objects and validation constraints. Your domain’s data model becomes the agent’s contract.
• RLM (Recursive Language Model) execution lets the agent write code to inspect large datasets, filter documents, compute aggregates and ask focused semantic questions through llmQuery(). Context lives in a persistent runtime, with intermediate results reusable across turns.
• Typed domain tools, dynamic tool discovery and specialist child agents connect reasoning to your actual systems. Skills and memory retrieval bring in relevant procedures and prior knowledge on demand.
• GEPA-backed agent.optimize() tunes instructions and demonstrations against your domain evaluations. You can score tool selection, delegation, forbidden actions and answer quality using your own definition of success.
• ACE-backed playbooks accumulate reusable task knowledge. Verified evolution can turn recurring failures into rules, accepting changes only when training performance improves and held-out evaluation stays within the configured regression tolerance.
That combination matters: explicit contracts, programmable context exploration, domain capabilities and measurable adaptation—all in one agent.
2
0
0
341
thanks for sharing all your wisdom, knowledge, skills and tools with us! So happy I stumbled upon you, @MaximeRivest with aiconvo.
1
1
2
357
tryingEveryThing retweeted
@ValerioCapraro To steal, one must take something that was not voluntarily given. These researches should state whether they opted out of model training. If so, that is a potentially big problem for OpenAI. If they did not, we can sympathize about the credit being lost but it is not stealing.
2
1
7
1,048
might be something for agnt @NathanWilbanks_ ?
100% this
This is why we built a WebAssembly- and isolates-based sandbox as a library that's Linux-compatible / POSIX-compliant
Vastly more resource efficient & lets you use more types of hardware
github.com/rivet-dev/agen…
1
0
2
151
can you please expand to europe, especially germany? ;)
told codex to design me a welded chair made of tubes and sheet metal from RMFG
it designed this step file, annotated all the welds, checked the DFM rules throughout the process, uploaded the design and got a quote, all from the CLI.
sent me a link to view it in
0
0
1
68
So I just had a chat about dspys flex and ax. Apparently it is not ported yet. Also there looks like a bit more to do.
@dosco if you are interested in the ai analysis: chatgpt.com/share/6a9e6851…
@dosco if you are interested in the ai analysis: chatgpt.com/share/6a9e6851…
1
0
1
235
have you seen this github.com/githubnext/cho… @MaximeRivest there might be some interesting UX behavior to adopt for github.com/MaximeRivest/a…
1
0
0
152
look at this eink awesomeness: store.bigme.vip/pages/bigme-hi…
Just preordered. hopefully @BigmeEInkTablet will deliver
Just preordered. hopefully @BigmeEInkTablet will deliver
0
0
2
99
Maybe reconsider given this: x.com/shiridesu/stat…
@huacnlee @zeddotdev this is why i cancel my plan on making gpui derived ui framework
1
0
1
75
I was introduced to your work by @FrostForger - so thanks to you both!
Almost 10% of the way to our goal of 500!
Once we hit 500 I’ll reinvest proceeds to start work on larger models like Qwen3.8-Flash-Next and quantize for all viable setups.
I’m humbled by your support, thank you to each of you who have subscribed!
patreon.com/AeonForge7
0
0
2
75
Hire this guy, you will not regret it @jihanyang13
@jihanyang13 I am very well suited to this role if professional experience and success equate to a degree. did game dev and character art for a long time before Ai research and now I do the AI stuff full time. Bridge both worlds with expert knowledge. Sounds really cool
1
0
1
71