J. Techie

💻 Entusiasta de la tecnología: dando vida al código desde '78.

[ComfyUI @ComfyUI]

Creator seungho__yeo ( IG ) built an entire anime-inspired 2D motion sequence - storyboard to final render in one hour. Inside Comfy.

seungho__yeo created this entirely in a single workflow pipeline:

→ Poster art
→ Storyboard frames
→ Transitions + animation
→ Edit + final render

What usually takes multiple apps was built in one tool.

[Arc Institute @arcinstitute]

The Proto team from @BrianHie's lab will be at the re:AGENT hackathon, a two-day weekend (8/15–8/16) in San Francisco for people interested in building AI scientists, datasets, or pipelines for biological design.

Apply by Thursday, July 30th, 12 PM PT at https://t.co/H2ddXzQBLP

[Juan Pablo Galeano R. @juangaleanorey]

YouTube Prueba Nuevas Herramientas para Crear Miniaturas Personalizadas con IA
https://t.co/tFPU6NPFHn

[Imbue @imbue_ai]

when i stopped having fun scrolling, graphed

(@Millanphilipose and Cindy Wei's new project to analyze happiness via face data while you scroll, using the front facing camera. apps like bouncer can alert you when you're not having fun anymore!)

[GitHub Projects Community @GithubProjects]

Control FreeCAD from Claude Desktop - turn natural language into 3D parts.

Explore it here:
https://t.co/TGJPRm1kyc

[Wildminder @wildmindai]

LTX2.3 Motion Enhancer LoRA
a general-purpose motion enhancer.
boosts fluidity and coherence for I2V.
works best stacked with other LoRAs at 0.65, or standalone 0.75–1.0.

https://t.co/tBcEU3Ga4H

[Ben Dickson @bendee983]

I used to write code by hand before Stack Overflow. Never copy-pasted tutorial files when I was learning something new. I would write again and again until I memorized syntax (to be fair, I had a lot of time on my hand at the time).

I continued to write code after Stack Overflow came around. Even if I found the solution there, I would not copy-paste anything that I hadn't done before.

It's the same with AI. You can take the copy-paste shortcut and just let the tool do the thinking and the doing for you, or you can build habits and muscle memory, get ideas and potential solutions from the AI, implement them by hand, learn new things, and then use AI as an amplifier.

There is always a choice.

[0xSero @0xSero]

@Bloomberg should write about how their name is being used to target Open Source devs with phishing attacks for the last 5 months.

I'm happy to share the entire dossier I have on these people and the patterns.

LMK

[0xSero @0xSero]

Hacking in the age of AI is quite stupid, you can't not leave trails, and now you have infinitely patient hunt dogs ready to chase you to hell.

Seeing this today I am way less doomery about cyber security, Codex / GLM were able to identify the trail back to December.

[Rohan Paul @rohanpaul_ai]

Ilya Sutskever: Humans aren’t AGI; they know little and learn constantly.

Pretrained AGI on the other hand, overshoots by skipping that trial-and-error phase. Real intelligence comes from ongoing learning, not from starting out “finished.”

[Rohan Paul @rohanpaul_ai]

Same acceleration, no news cycle, no panic.

Software solved this already with canary rollouts, shadow traffic and instant rollback.

Current model weights are a deployment artifact with far worse observability

[Chris @ChrisGPT]

Dario Amodei just responded to everyone accusing Anthropic of wanting to ban open weight models to protect its own business

“Anthropic has never advocated for a ban on open-weights models.”

He actually goes even further and says:

“Open-weights models that don’t have dangerous capabilities are a public good: they don’t cost anything besides the compute needed to run them, and they provide value to businesses, developers, and researchers.”

His argument is basically that banning US companies from using Chinese open-weight models would do almost nothing to address the national security threats he is actually worried about.

He notes that the most dangerous Chinese model probably would not be released publicly in the first place -

“In fact, the most dangerous model may be one that is trained in secret and handed only to the People’s Liberation Army for use in drones and the Ministry of State Security for surveillance and repression.”

He also is not pretending that open-weight models have zero additional risk.

TL;DR - Once sufficiently capable weights are released, the guardrails can be removed, usage cannot be monitored, and the model can never truly be withdrawn. But banning legitimate US companies from using them still would not stop bad actors.

[chuplung @choopyplug1]

Dario Amodei (CEO, Anthropic):

"every software engineer just got a deadline. 6 to 12 months until models do everything they do. end to end"

engineers at Anthropic don't write code anymore. they edit what the model writes

"if I had to guess, this goes faster than people imagine"

the thing both he and Demis Hassabis agreed to watch: AI systems building AI systems. that's the loop that changes everything

~30 minutes with the man building the future

save this. then read the article below ↓

[Turing Post @TheTuringPost]

Recursive self-improvement (RSI) usually refers to systems improving the machinery that makes it intelligent.

AREX family of RSI deep research agents applies this term much more narrowly

▪️ Its weights stay fixed during a run while its research state is repeatedly rewritten.

Here's how it works:

- The agent searches and drafts an answer (inner research loop).
- A second loop (outer self-improvement loop) audits the answer against its constraints, preserves verified, marks unresolved claims, and decides whether to refine or restart.
- A learned context tool compresses the research history, also keeping rejected candidates and the updated research plan.

The results show real profit.
On BrowseComp, context updating raised accuracy from 59.6% to 71.4%. Adding the outer loop took it to 82.5%.

▪️ The practical contribution is that each round of verification shapes the next round of research.

So this isn't a model teaching itself to become smarter. It's a system making its own research process better, one iteration at a time.

But there is one thing that is not obvious from the study. AREX allows up to 300 research turns and 5 outer loops, although the paper doesn't compare those gains against the additional compute cost.

[arc. @arceyul]

Genspark Design is seriously impressive.

I tested it on a fictional amusement park brand “ThrillVille" to see how far one visual direction could go.

Poster, layered edits, website, brand video

all from one direction @genspark_ai

[Timothy Chen @tnachen]

Have always wanted to redo our @EssenceVenture website and finally decided it's time, also decided it's time to move off Squarespace as well.

So used @variantui to find a good starting point on the design side, and built / iterated the whole website using @claudeai and have it deployed the website with custom DNS all working straight in Claude with @insforge

https://t.co/dOQVUTzxVv

[Berkay @KBerkayglc]

Hello, hello! I am on trying the prototype. 🥳

#indiedev #gamedev #LookUp

[Synopsis @abhinavflac]

🚨 opus 5 one-shotted this whole thing

real time bayer dither in @threejs
it wrote its own glb writer and packs the gltf binary container

[Chema Alonso @chemaalonso]

El lado del mal - Cómo desplegar Zero Trust para Agentes IA en Cloudflare (1) https://t.co/p46JRmd5gs #Cloudflare #IA #AI #InteligenciaArtificial #AgentesIA #AgenticAI #ZeroTrust #Claude #Anthropic #hardening

[Daniel Eth (yes, Eth is my actual last name) @daniel_271828]

Of course, let’s also be clear-eyed here about the risks. Plenty of classic stories about genies granting poorly-specified wishes or otherwise scheming against their humans