@3CatInfo Vinga, som-hi altre cop. Però eh, rodalies ja funciona amb normalitat, les obres dels túnels ja han acabat, tot bé! Ah, i tots a pagar el bitllet eh! I revisor per fotreli pal a qui no pagui en motiu de protesta per la merda de servei que seguim tenint
Es van comprometre i ens van assegurar que durant el tall tots els trens de la R2 Sud serien llargs.
Doncs avui, un altre cop trens curts. És intolerable que la gent es quedi a terra o hagi de viatjar en aquestes condicions.
Què explica avui el que no arriba a treballar?
Anthropic pays $750K/ year per senior engineer.
The creator of Claude Code just revealed his coding setup at the Sequoia AI session.
Boris Cherny: "100% of my code is written by Claude Code. I run around 100 agents at one time."
free. 24 minutes. watch it
then read article below
"AI is approaching human-level reasoning." ARC Prize just proved it isn't.
> Humans: 100%.
> GPT-5.4: 0.26%.
> Claude Opus 4.6: 0.25%.
> Grok-4: 0.00%.
The benchmark gives agents no instructions, no goal description, no rules.
Figure it out. Humans do. Frontier AI cannot.
ARC-AGI-1 tested pattern recognition.
ARC-AGI-2 tested multi-step reasoning.
Both got saturated.
ARC-AGI-3 tests something current AI fundamentally cannot do.
The agent is dropped into a novel interactive environment with zero context. No goal. No rules. No explanation of what winning even means.
It has to explore, infer the win condition, build a model of how the environment works, and execute a plan all from scratch, on first contact.
Humans walk in cold and solve these in under 10 minutes.
Frontier models score below 1%. The benchmark was designed specifically to prevent the shortcuts that broke ARC-AGI-1 and 2.
Gemini 3 Deep Think was caught mid-reasoning using the exact integer-to-color mapping from ARC-AGI tasks without being told what the task was.
The training data had leaked in. ARC-AGI-3 inverts the public-to-private dataset ratio, keeps the private set out-of-distribution from everything public, and scores on action efficiency rather than binary pass/fail.
You don't just have to solve it.
You have to solve it as efficiently as a human would.
The harness result is the most revealing number in the paper.
With a specifically engineered harness built for known environments, Opus 4.6 scores 97.1% on one public environment.
On a different environment the harness wasn't built for: 0.0%. The model can solve these environments when a human has pre-loaded the exact strategy. Without that scaffolding it scores 0.25% overall.
The capability exists. The general intelligence walking into something new and figuring it out does not.
The full scoreboard at launch:
→ Humans: 100% solve rate, median 7.4 minutes on first contact, no instructions
→ Gemini 3.1 Pro Preview: 0.37%
→ GPT-5.4 High: 0.26%
→ Claude Opus 4.6 Max: 0.25%
→ Grok-4 Beta Reasoning: 0.00%
→ Harness-assisted Opus 4.6 on seen environment: 97.1% drops to 0.0% on unseen environment
→ Random policy cannot solve non-tutorial levels more than 1 in 10,000 attempts by design
→ ARC-AGI-3 is the only unsaturated general agentic intelligence benchmark as of March 2026
The gap exposes something precise about what current AI is and isn't.
LRMs excel when the domain has training coverage and a verifiable correctness signal. ARC-AGI-3 has neither. Every environment is novel.
There is no verifier. The model has to discover the win condition before it can attempt to win. Human reasoning is not bound by domain knowledge.
Current AI reasoning still is. That is the gap ARC-AGI-3 is measuring and right now it is almost total.
ARC-AGI-1 called the reasoning era before it arrived. ARC-AGI-2 measured how far it had come.
ARC-AGI-3 just showed where the ceiling actually is. Every company is racing to claim AGI.
The benchmark that has been right every time says we are not close.
Below 1% is where we are. 100% is where humans are.
Con més sabem, més consternats i enfadats
Ja no confiem en polítics i lobbys
Ens heu fet tant de mal
Estan matant rodalies tenim la sospita k es premeditat es imposible ser tant incompetents i tenir tan poca vergonya
Fora renfe i adif de Catalunya
https://t.co/TbVaojy1sx
🙄El país és un acudit, per anar del 🚢port de Barcelona a Port Bou ara les mercaderies han de passar per Lleida i fer 330km extra.
👉Tot per no mantenir un túnel de 900 m.
Això amics i amigues és Adif, és Óscar Puente, és Espanya.🥳🥳
Mientras tanto en China 🇨🇳
🇨🇳Un robot de mantenimiento ferroviario que repara automáticamente fisuras en la superficie de los rieles a una velocidad máxima de 10,000 RPM y detecta figuras de tan sólo o,1 mm ya en funcionamiento.
La voz de los más penalizados por el corte ferroviario del Garraf: "He decidido huir del tren". La R-2 sur y los usuarios de Regionals que viven más allá de Vila-seca están viviendo un auténtico calvario por el pésimo servicio ferroviario
https://t.co/wumpSOjum5
Túnel de Pajares (Astúries): 3000 viatgers diaris amb una inversió de 4 mil milions d'euros
Túnel ferroviari del Garraf (Catalunya): 40.000 passatgers diaris, 30 anys de desinversió crònica.
Ho enteneu ara, oi?
El @diariARA explica en una infografia la importància d'aquesta obra: els @regsud de Tarragona, Reus, Tortosa i Valls podrien anar per les vies de fora sense que els @rod2sudcat els entorpissin
Imatge bastant ilustrativa de la situació lamentable que vivim els catalans. Una xarxa ferroviària que cau a trossos. Una situació insostenible, que ens afecta a tots els qui hem de moure'ns per treballar i ens està afectant física i psicològicament. I cada cop va a pitjor...
Mapa lineas de CAT cortadas a día de hoy. Impresionante.
Como impresionante es que ADIF envie a CAT a 400 operarios de repente, en lugar de hacer mantenimiento preventivo como FGC lleva haciendo hace años. Cómo @Adif_es pudo desmontar sus bases en CAT sin que gobierno 🇪🇸 actuase?
Agafo un tren amb retard de la R2 al Clot, ple a rebentar. Una noia em diu que l’ha agafat a Bellvitge i que els han fet canviar de comboi a Sants perquè el maquinista no s’ha presentat. Que ahir va passar el mateix. Tenim el país segrestat per fanàtics i governat per ineptes.