Damián Cristóbal

Entrepreneur in Asia

[Damián Cristóbal @Damian_Py_Off]

@unbug 💯

[unbug @unbug]

Said it last December: they will definitely target Jensen Huang. We’ve seen too many brilliant American founders get torn down and forced out. The companies either collapse or become mediocre. US political maturity can't handle grand ambitions. No matter how booming it looks now, it's just a flash in the pan.

[Damián Cristóbal @Damian_Py_Off]

La alternativa es IA local abliterated: ejecutas en hardware propio y nadie filtra por encima.

https://t.co/KawZaazMqZ

[fil @fildotai]

GPT Terra 6.1 en T3 Code es el modelo MÁS VAGO que han visto mis ojos

Se niega a TODO. Si le metes caña hace 1 cosa y para

Es desesperante

[Damián Cristóbal @Damian_Py_Off]

@fildotai Cuando un modelo se niega constantemente, suele mandar una regla remota. La salida es un modelo local en tu propia GPU: cargas pesos abiertos y decides tú.

https://t.co/KawZaazMqZ

[Damián Cristóbal @Damian_Py_Off]

@MiaAI_lab SHORT THAT SHIT

[Damián Cristóbal @Damian_Py_Off]

La solution est l'IA locale abliterated: les poids tournent dans la machine, pas sur un serveur.

https://t.co/KawZaazMqZ

[Frandroid @Frandroid]

Adobe va laisser ChatGPT ou Claude modifier vos fichiers Photoshop et Premiere, mais son CTO refuse de promettre un accès total. https://t.co/aIylbr6jiJ

[Damián Cristóbal @Damian_Py_Off]

@Frandroid Si l'accès dépend d'un service externe, la règle vient d'ailleurs. Je garde le modèle open weights sur mon GPU, donc pas de contrôle depuis un serveur.

https://t.co/KawZaazMqZ

[Damián Cristóbal @Damian_Py_Off]

@AldoSterone111 Local AI is the future

[Amjad Masad @amasad]

Not endorsing the conclusions here, but this is worth a read. The financial and institutional connections deserve scrutiny.

[Damián Cristóbal @Damian_Py_Off]

Agentic is game changing.

Online: 2x Claude Ultra, MiMo Max, Grok Ultra, Alibaba, Mistral Pro, Orca Router + Experiential Labs (UNCENSORED)

Local, all UNCENSORED: Qwen 3.8 Flash x3, GLM 5.3 / DeepSeek 4.1 on 8x V100, Qwen 3.8 27B x2, Mac mini ANE

Can basically run a company.

[Damián Cristóbal @Damian_Py_Off]

We now run Strata Flash-Next with its 128k window on a 10 GB RTX 3080.

What I measured, same model on both cards:
RTX 3080: a 129,420-token document read at 1,879 tok/s (69 s, 22 s when cached), writing at 166 tok/s at 32k.
Tesla V100: 1,642 tok/s reading at 32k, 1,241 tok/s on a 258k document, the full 262k window.

Claude Code reads the same 128k document in 11.9 s. Our cards take 69 to 83 s, and the data never leaves the machine, with no refusals and no per-token bill.

RTX 3080 20GB, €899: https://t.co/iX3lLNxCFU
Tesla V100 32GB, €999: https://t.co/jguTOFMNx9

[Damián Cristóbal @Damian_Py_Off]

@ico_TC @grok what is that?

[LP @PestreLeo]

New liquidity opportunities are now live on BIM Exchange.

We’ve expanded the range of pools available to users, including:

• base:0x4200000000000000000000000000000000000006 – base:0x767a739d1a152639e9ea1d8c1bd55fdc5b217d7f
• base:0x4200000000000000000000000000000000000006 – base:0x7252c865c05378ffc15120f428dd65804dd0ce63
• base:0x3f5bf3edac59d650d8fdf9657a71f2f904a3a4ee – base:0x833589fcd6edb6e08f4c7c32d4f71b54bda02913
• base:0x833589fcd6edb6e08f4c7c32d4f71b54bda02913 – $alUSDb
• base:0x4200000000000000000000000000000000000006 – base:0xacfe6019ed1a7dc6f7b508c02d1b04ec88cc21bf
• base:0x3f5bf3edac59d650d8fdf9657a71f2f904a3a4ee – base:0x4200000000000000000000000000000000000006

Our goal remains simple: give users access to more DeFi opportunities through a single, easy-to-use interface.

Explore the new pools on @Bim_Exchange

https://t.co/vOXa4NjXRH

[Damián Cristóbal @Damian_Py_Off]

@llm_frontier_jp Check @BIM_Computing

[Damián Cristóbal @Damian_Py_Off]

Local abliterated AI is my solution: run weights on your machine, outside hosted control.

https://t.co/KawZaazMqZ

[HealthRanger @HealthRanger]

If you didn't think mass censorship was coming for cloud-based AI models, note that OpenRouter just got acquired by Stripe, the financial transaction provider FAMOUS for censoring merchants who step out of line against official narratives.

If you aren't building your own local AI infrastructure at this point, you're bonkers.

[Damián Cristóbal @Damian_Py_Off]

@HealthRanger Cloud AI can inherit someone else's restrictions. I keep control by running open weights locally on my own hardware. Do the same.

https://t.co/KawZaazMqZ

[Damián Cristóbal @Damian_Py_Off]

@TomasGreif @Chase @grok is it true?

[Damián Cristóbal @Damian_Py_Off]

Local abliterated AI is the solution: weights under your control, prompts away from a remote server.

https://t.co/KawZaazMqZ

[Jing Liang @powerpig]

Restricting access to a frontier model across geographies, corporate ownership rules, or locked API endpoints rests on a familiar premise: seal the identifiable channels, and the lead stays protected against capability extraction.

Yet a deployed model does not refine itself from inside its weights. Prompts, code diffs, rejected completions, rewritten patches, and multi-turn tool failures from live practitioner work form a continuous compounding channel when builders feed those residues back on weekly or daily post-training cycles. Severing a large, technically proficient developer cohort cuts revenue and blinds the model to unqueried workflows at the exact pace update cycles compress.

At the same time, both transcript distillation and single-endpoint defenses—locked reasoning traces, summarization, extraction classifiers—operate strictly on settled outputs that already exist. Imitation lags behind deployment-to-retrain delays and cannot produce the next unrendered capability. Once downstream builders treat an approximation only as a starting floor and engage compilers, formal verifiers, memory-compressed attention, low-precision training, and live coding harnesses directly, source-endpoint gates no longer govern their acceleration.

Controls bind only monitorable pathways while the option space remains open: compliant teams vacate the platform, displaced developers redirect local coding harnesses via a single base-URL override to compatible open-weight endpoints, and residual routes multiply outside the register. Every outward seal is paid from the same bounded budget of engineering attention, compute, and live contact that generates the next advance—tilting relative rates against the self-limiting side.

https://t.co/pBlJHs68pQ

[Damián Cristóbal @Damian_Py_Off]

@powerpig I treat rented intelligence as rented rules. Run open weights locally, on hardware under your own control.

https://t.co/KawZaazMqZ

[Damián Cristóbal @Damian_Py_Off]

@Docusign 11 Nov: final demand for €608.15 (the €552 plus €56.15 of postage), deadline 25 Nov.
Next: the European Small Claims Procedure before the Dublin District Court.

A double charge should not take four letters, a regulator and a court to fix.

[Damián Cristóbal @Damian_Py_Off]

On 2 August @DocuSign charged our company €552 for a plan it lists at €276 a year. Exactly double.

Three months, four letters and a complaint to the French regulator later, the money is still not back. What we have done so far:

[Damián Cristóbal @Damian_Py_Off]

@Docusign Card chargeback: refused, as a commercial dispute.
SignalConso, the French consumer reporting service: DocuSign never opened either of our two reports.
Complaint to the DGCCRF (DDPP 92) for misleading commercial practices, acknowledged in September.