STEM

[Marv]

Apple ha sacado un paper para decir que los roboces no razonan.


https://machinelearning.apple.com/research/gsm-symbolic
https://arxiv.org/abs/2410.05229

We hypothesize that this decline is because current LLMs cannot perform genuine logical reasoning; they replicate reasoning steps from their training data. Adding a single clause that seems relevant to the question causes significant performance drops (up to 65%) across all state-of-the-art models, even though the clause doesn't contribute to the reasoning chain needed for the final answer.

O sea, que sólo replican los pasos de razonamiento que están en su entrenamiento.

Había una manera más sencilla de decirlo: estos bichos no entienden nada de lo que están haciendo, porque no pueden entenderlo. Son estructuras matemáticas implementadas en hardware. Autocompletadores on steroids.

Por supuesto, ustedes ya lo sabían, porque se dijo aquí.

Vivimos tiempos curisosos, donde hace falta montar una burrada de estudio para desmontar a los vendedores de crecepelo, porque la gente es genuinamente engañada y están indefensos. Incluso los que se supone que deberían protegernos del misticismo. Tecnosocurantismo.

No es por ser tocahuevos, pero son estructuras matemáticas implementadas en un software, y para ello se valen de un hardware. El hardware es un mero sustrato físico para que el software pueda operar.

Si hay algo basado exclusivamente en hardware somos nosotros, aunque filosóficamente podría ser discutible según se enfoque. Las calculadoras antiguas electrónicas sí que eran puro hardware y no podían cambiar sus funciones ni eran programables (que es lo que da pie al concepto de software).

No tiene relevancia para la discusión, pero era por no liar terminología.

No tengo ningún problema con la matización, porque efectivamente es irrelevante. Sin embargo disiento.

La estructura matemática subyacente son los hipergrafos ponderados.
El harware son las redes neuronales artificiales.
Y el software es la secuencia de instrucciones que ejecutada, explota el hardware.

El software en sentido filosófico no es la implementación, sino la función, aunque sé perfectamente que la industria del software llama implementar a escribir código fuente. Desde mi punto de vista, no es correcto.

Con respecto a lo que quiere Apple, yo no lo sé pero por supuesto que algo tendrá que ver con aguar la fiesta del hype. Incluso puede tener intereses directos en ello, porque es partner de OpenAI y entiendo que querrá que no se suban a las barbas con los precios.

El hardware no pueden ser redes neuronales, porque tal concepto es puramente abstracto y existe sólo en el software residiendo en la memoria en forma de programa+datos exactamente igual que Excel o Chrome. Y cuando ejecutamos Excel, ¿a que no lo llamamos neuronas artificiales ni lo llamamos nada especial? Pues eso.

Ni el hardware ni su nombre ni su concepto cambian en función de lo que ejecutes en él. No hay mucha discusión posible. Si quieres llámalo Mariano, pero el hardware es lo que es.

El hardware sigue siendo una CPU+GPU y memoria, tu planteamiento no tiene sentido. Sólo en los casos en los que se ejecuten las operaciones en hardware diseñado específicamente para ejercer esa función de neuronas artificiales (que por otro lado no tienen nada de especial, son objetos matemáticos como podrían ser una matriz o un quaternion) te podría dar la razón.

No me hice entender. No quería decir que las neuronas artificiales existen como bichitos de hardware mecánicos dentro de los chips, hablándose entre sí y remedando lo que ocurre en nuestras cabezas.

Lo que digo es que el software no implementa nada, a nivel filosófico (y la filsofía la habías traido tú).
Escribe el cálculo de Pi en python sobre una servilleta y déjala sobre la mesa. Ahí no hay implementación*. El software son reglas.

[Si llamas software al fichero ejecutable, te diría que es lo mismo, pero en un lenguaje más cercano a las máquinas. Si llamas software a lo que hay en la memoria principal en un momento dado de ejecución (con el conjunto de instrucciones máquina + los datos temporales de ejecución), te vuelvo a decir que es lo mismo, sólo que ha cambiado de sitio. Y si propones que esto último es físico, entonces es hardware.]

Como símil pongo la implementación de una medida política (nueva función), que se desarrolla en su reglamento (software). Esa medida llegará a estar implementada cuando se ejecute en el mundo real, en un entorno de 'funcionarios haciendo cosas' (hardware). No antes. El reglamento per se no implementa nada, se puede guardar en un cajón y no implementarlo nunca.

Si queremos abstraer, los LLM son un conjunto de números (que llaman parámetros) estáticos, que han quedado establecidos tras el entrenamiento. Sin las operaciones de pregunta + matcheo + respuesta no son gran cosa; han triunfado en el momento de añadir los transformers, que permiten el autocompletado en bucle y la sensación de inteligencia. El proceso de autocompletado es lo que realmente da 'anima' a la cosa; pero eso no es la red neuronal, es el software que la aprovecha (a eso me referíaa con la función).

Resumiendo, es cierto que la implementación no es una correspondencia 1:1 con ninguno de los elementos que intervienen, sino que está distribuida en todos ellos: datos, software y hardware, con una línea difícil de trazar, y eso es lo filosófico aquí. ¿No es el hardware solamente? Concedido. Pero tampoco es el software solamente.

A nivel práctico, no cambia nada.

---

(*) Sobre la nomenclatura.
RAE: Implementar: poner en funcionamiento, aplicar métodos, medidas, etc., para llevar algo a cabo

Me quedo con "aplicar métodos".

El sujeto implícito (hardware), aplica (ejecuta) métodos (software)

[Cadavre Exquis]

https://cincodias.elpais.com/fortunas/2024-12-01/gonzalo-abellan-matteco-hacemos-que-metales-no-preciosos-se-comporten-como-si-lo-fueran.html

Saludos.

[Cadavre Exquis]

TED | How AI Is Saving Billions of Years of Human Research Time | Max Jaderberg | TED

Saludos.

[Cadavre Exquis]

AI Boosts Materials Discovery By 44% at Major US Lab
Posted by msmash on Monday December 09, 2024 @10:08AM from the already-paying-dividends dept.

AI-powered teams at a major U.S. materials company discovered 44% more new materials and filed 39% more patents compared to teams using standard methods, according to a study by MIT economist Aidan Toner-Rodgers.

The research, conducted at an unnamed corporate laboratory with over 1,000 scientists, tracked the implementation of a custom machine-learning system combining graph neural networks with reinforcement learning. The AI tool, deployed in 2022, was pre-trained on crystal and molecular structure databases. Top-performing scientists showed the greatest gains with AI assistance, while lower-ranked researchers saw minimal benefits. The AI-designed materials demonstrated higher novelty compared to human designs, based on patent text analysis.

The company's secrecy limits independent verification of results, according to University College London chemist Robert Palgrave. Researchers using AI reported lower job satisfaction, citing reduced creative involvement in the discovery process.

Saludos.

[Cadavre Exquis]

A new era for batteries: Argonne leads $50M sodium-ion innovation push
Six national laboratories and eight universities will seek to increase the supply diversity of electric vehicle batteries and relieve supply chain concerns

2024.11.21

Argonne Distinguished Fellow Christopher Johnson in the lab working on advanced batteries for electric vehicles and other applications. (Image by Argonne National Laboratory.)

The U.S. Department of Energy (DOE) has awarded $50 million over the next five years to establish the Low-cost Earth-abundant Na-ion Storage (LENS) consortium. Led by DOE’s Argonne National Laboratory, the consortium includes DOE’s Brookhaven National Laboratory, Lawrence Berkeley National Laboratory, Pacific Northwest National Laboratory, Sandia National Laboratories and SLAC National Accelerator Laboratory.

The LENS consortium aims to develop high-energy, long-lasting sodium-ion batteries using safe, abundant and inexpensive materials. This initiative addresses a critical need to reduce U.S. dependence on the limited and strategically important elements used in lithium-ion batteries, paving the way for a more sustainable future in electric-vehicle technology.

“By leading the LENS consortium, Argonne will push sodium-ion battery technology forward and contribute to a secure energy future for everyone,” said Argonne Director Paul Kearns. ​“Our scientific expertise and dynamic collaborations in this important field will strengthen U.S. competitiveness.”

“The challenge ahead is improving sodium-ion energy density so that it first matches and then exceeds that of phosphate-based lithium-ion batteries while minimizing and eliminating the use of all critical elements.” — Venkat Srinivasan, LENS consortium director

At present, lithium-ion batteries dominate the global energy storage market for both vehicles and stationary storage. They power devices ranging from smartphones to electric vehicles and can store energy from renewable sources like solar and wind.

Relying on any single battery chemistry, however, creates vulnerabilities, and the dominant batteries today include the critical elements of lithium, cobalt and nickel. Sodium, as an abundant element, can reduce risk and increase supply chain resilience by providing a wider variety of cost-effective options.

The U.S. is particularly well-suited to supply both the raw materials and innovation for sodium-ion technology because the country produces a substantial amount of the world’s sodium chloride (table salt) and sodium. Sodium-ion batteries have the potential to eliminate not just lithium in some applications, but also cobalt and nickel, providing a more affordable and sustainable solution.

However, sodium-ion batteries store less energy per unit weight and volume, which results in a lower driving range — a stumbling block to such batteries competing with lithium-ion batteries.

“The challenge ahead is improving sodium-ion energy density so that it first matches and then exceeds that of phosphate-based lithium-ion batteries while minimizing and eliminating the use of all critical elements,” said Venkat Srinivasan, director of the LENS consortium and of the Argonne Collaborative Center for Energy Storage Science (ACCESS). ​“Importantly, any improvements must not compromise other performance metrics such as cycle life and safety.”

“Sustaining the deployment of electric vehicles requires affordable and sustainable battery chemistries,” said Will Chueh, director of the SLAC-Stanford Battery Center. ​“The bold technical targets of LENS aim to transform sodium-ion batteries from a promising technology to a viable component of tomorrow’s electric vehicles.”

To achieve this goal, Argonne has convened a world-class team of researchers from national laboratories and universities. Each participant brings deep experience studying sodium-ion batteries. Collectively, they will work to discover and develop high-energy electrode materials, improve electrolytes, and design, integrate and benchmark battery cells.

“Sodium-ion batteries can play an important role in society’s need for inexpensive energy storage,” said Gerd Ceder, a senior faculty scientist in Berkeley Lab’s Materials Sciences Division. ​“Foundational research in the area of new materials discovery, advanced manufacturing, and characterization is critical to the development and deployment of competitive sodium-ion battery technology. Berkeley Lab is proud to be part of this consortium.”

An advisory board comprising well-established and emerging companies will provide the consortium with valuable industry perspectives, with a goal of nurturing a U.S. ecosystem for sodium-ion batteries.

LENS will be part of a growing portfolio within DOE on sodium-ion batteries, which includes research into the use of this emerging chemistry in electric vehicle and grid storage applications.

The consortium includes eight university partners: Florida State University, University of California San Diego, University of Houston, University of Illinois Chicago, University of Maryland, University of Rhode Island, University of Wisconsin–Madison and Virginia Tech. With the involvement of all 14 partners, LENS will play a key role in training a new generation of battery scientists and researchers.

“Our world is on the verge of a profound shift in how we power our everyday lives,” said Feng Lin, professor of chemistry at Virginia Tech. ​“With the combined expertise of the LENS consortium, we now have a unique opportunity to pioneer new battery technologies for electric vehicles and to train a new generation of scientists and engineers who will contribute to our domestic battery innovation and manufacturing.”

LENS is supported by DOE’s Office of Energy Efficiency and Renewable Energy’s Vehicle Technologies Office.


Argonne National Laboratory seeks solutions to pressing national problems in science and technology by conducting leading-edge basic and applied research in virtually every scientific discipline. Argonne is managed by UChicago Argonne, LLC for the U.S. Department of Energy’s Office of Science.

The U.S. Department of Energy’s Office of Science is the single largest supporter of basic research in the physical sciences in the United States and is working to address some of the most pressing challenges of our time. For more information, visit https://​ener​gy​.gov/​s​c​ience.

Saludos.

[Cadavre Exquis]

New Magnetic Flow Has Potential To Revolutionise Electronic Devices
Posted by msmash on Wednesday December 11, 2024 @01:02PM from the moving-forward dept.

An international research team has for the first time imaged and controlled a type of magnetic flow called altermagnetism, which physicists say could be used to develop faster and more reliable electronic devices. Financial Times:

A groundbreaking experiment at a powerful X-ray microscope in Sweden provides direct proof of the existence of altermagnetism, according to a paper published in Nature on Wednesday. Altermagnetic materials can sustain magnetic activity without themselves being magnetic.

The team from the UK's Nottingham university that led the research said the discovery has revolutionary potential for the electronics industry. "Altermagnets have the potential to lead to a thousand-fold increase in the speed of microelectronic components and digital memory, while being more robust and energy-efficient," said senior author Peter Wadley, Royal Society research fellow at Nottingham.

Hard disks and other components underpinning the modern computers industry process data in ferromagnetic materials, whose intrinsic magnetism limits their speed and packing density. Using altermagnetic materials will allow current to flow in non-magnetic products.

Saludos.

[Cadavre Exquis]

Leading Scientists Urge Ban On Developing 'Mirror-Image' Bacteria
Posted by BeauHD on Thursday December 12, 2024 @10:30PM from the cease-and-desist dept.

An anonymous reader quotes a report from Science Magazine:

It would be a nightmare lab leak: Synthetic bacteria escape the petri dish and unleash a global plague that life on Earth is incapable of defending itself against. That's the concern raised by a group of eminent researchers in a Policy Forum published online today in Science. The commentary's 38 authors, from a broad range of disciplines, argue that governments worldwide should prohibit research and funding aimed at creating so-called mirror-image bacteria whose chemical makeup differs in a fundamental way from that of naturally existing organisms.

All of life's primary biomolecules can exist in two mirror-image forms, like a left and right hand. But only one form is found in nature. Proteins are left-handed, for example, and DNA and RNA are right-handed. Synthetic biologists have previously synthesized mirror-image proteins and genetic molecules. And mirror-image amino acids and peptides -- the building blocks of proteins -- have been incorporated into several approved drugs. Because natural enzymes struggle to break down mirror-image biomolecules, these components help the drugs survive longer in the body. [...]

The concern, he and others say, is that taking this line of work many steps further could result in fully mirror-image bacteria that could reproduce. Such organisms would likely be able to infect and potentially harm a wide range of microbes, plants, and animals while resisting the molecules that enable predators to kill and digest existing microbes. "They are essentially unassailable to those enzymes," says John Glass, a co-author and synthetic biologist at the J. Craig Venter Institute. Animals' immune systems would also struggle to cope with mirror bacteria. They "would be invisible to the immune system until it was too late," says Timothy Hand, a co-author and immunologist at the University of Pittsburgh. The Policy Forum authors acknowledge it will be at least a decade before synthetic biologists will be capable of creating these life forms. Nevertheless, they recommend halting all research aimed at that goal and urge funding agencies not to support it.

"It's hard to overstate how severe these risks could be," says Ruslan Medzhitov, an immunologist at Yale University and one of the authors. "If mirror bacteria were to spread through infected animals and plants, much of the planet's many environments could be contaminated. ... Any exposure to contaminated dust or soil could be fatal."

Jack Szostak, a co-author and a 2019 Nobel Prize-winning chemist at the University of Chicago, adds: "The result could be catastrophic irreversible damage, perhaps far worse than any challenge we've previously encountered."

Saludos.

[Cadavre Exquis]

Single Crystal Lithium-Ion Batteries Last 8x Longer, Researchers Show
Posted by EditorDavid on Sunday December 15, 2024 @11:34AM from the 20,000-cycles dept.

Researchers used Canada's national synchrotron light source facility "to analyze a new type of lithium-ion battery material — called a single-crystal electrode — that's been charging and discharging non-stop in a Halifax lab for more than six years," reports Tech Xplore.

The results? The battery material "lasted more than 20,000 cycles before it hit the 80% capacity cutoff," which they say is equivalent to driving 8 million kms (nearly 5 million miles). That's more than eight times the life of a regular lithium-ion battery that lasted 2,400 cycles before reaching the 80% cutoff — and "When the researchers looked at the single crystal electrode battery, they saw next to no evidence of this mechanical stress." (One says the material "looked very much like a brand-new cell."

Toby Bond [a senior scientist at the CLS, who conducted the research for his Ph.D.] attributes the near absence of degradation in the new style battery to the difference in the shape and behavior of the particles that make up the battery electrodes... The single crystal is, as its name implies, one big crystal: it's more like an ice cube. "If you have a snowball in one hand, and an ice cube in the other, it's a lot easier to crush the snowball," says Bond. "The ice cube is much more resistant to mechanical stress and strain." While researchers have for some time known that this new battery type resists the micro cracking that lithium-ion batteries are so susceptible to, this is the first time anyone has studied a cell that's been cycled for so long...

Bond says what's most exciting about the research is that it suggests we may be near the point where the battery is no longer the limiting component in an EV — as it may outlast the other parts of the car. The new batteries are already being produced commercially, says Bond, and their use should ramp up significantly within the next couple of years. "I think work like this just helps underscore how reliable they are, and it should help companies that are manufacturing and using these batteries to plan for the long term."

Thanks to long-time Slashdot reader schwit1 for sharing the news.

Saludos.

[Cadavre Exquis]

Microsoft Announces Phi-4 AI Model Optimized for Accuracy and Complex Reasoning
Posted by EditorDavid on Monday December 16, 2024 @12:34AM from the Bing-bot's-brother dept.

An anonymous reader shared this report from Computerworld:

Microsoft has announced Phi-4 — a new AI model with 14 billion parameters — designed for complex reasoning tasks, including mathematics. Phi-4 excels in areas such as STEM question-answering and advanced problem-solving, surpassing similar models in performance. Phi-4, part of the Phi small language models (SLMs), is currently available on Azure AI Foundry under the Microsoft Research License Agreement and will launch on Hugging Face [this] week, the company said in a blog post.

The company emphasized that Phi-4's design focuses on improving accuracy through enhanced training and data curation.... "Phi-4 outperforms comparable and even larger models on tasks like mathematical reasoning, thanks to a training process that combines synthetic datasets, curated organic data, and innovative post-training techniques," Microsoft said in its announcement. The model leverages a new training approach that integrates multi-agent prompting workflows and data-driven innovations to enhance its reasoning efficiency. The accompanying report highlights that Phi-4 balances size and performance, challenging the industry norm of prioritizing larger models... Phi-4 achieved a score of 80.4 on the MATH benchmark and has surpassed other systems in problem-solving and reasoning evaluations, according to the technical report accompanying the release. This makes it particularly appealing for domain-specific applications requiring precision, like scientific computation or advanced STEM problem-solving.

Microsoft emphasized its commitment to ethical AI development, integrating advanced safety measures into Phi-4. The model benefits from Azure AI Content Safety features such as prompt shields, protected material detection, and real-time application monitoring. These features, Microsoft explained, help users address risks like adversarial prompts and data security threats during AI deployment. The company also reiterated that Azure AI Foundry, the platform hosting Phi-4, offers tools to measure and mitigate AI risks. Developers using the platform can evaluate and improve their models through built-in metrics and custom safety evaluations, Microsoft added... With Phi-4, Microsoft continues to evolve its AI offerings while promoting responsible use through robust safeguards. Industry watchers will observe how this approach shapes adoption in critical fields where reasoning and security are paramount.

Saludos.

[pollo]

Driverless Bus Service In Scotland To Be Withdrawn Due To Lack of Interest

Posted by BeauHD on Tuesday December 17, 2024 @10:30PM from the falling-short-of-expectations dept.

An anonymous reader quotes a report from The Guardian:

The UK's first driverless bus service, originally heralded as a breakthrough of global significance, is being withdrawn from service because too few passengers used it. The autonomous buses, operated by Stagecoach, have been running between Fife and Edinburgh along a 14-mile route over the Forth road bridge since May 2023 to relieve the heavy congestion which can bring traffic to a standstill.
The CAVForth service, a collaboration between Fusion Processing, the coach-building company Alexander Dennis, Napier University in Edinburgh and the Bristol Robotics Lab -- a joint venture between the University of Bristol and the University of the West of England -- was touted as the most ambitious and complex in the world. Built at an estimated cost of more than 6 million pounds, partly funded by the UK government, the fleet of five single-decker buses had the capacity to carry 10,000 passengers a week but needed two crew on board for safety reasons. Stagecoach said in a statement: "We are proud to have achieved a world first with our CAVForth autonomous bus service, demonstrating the potential for self-driving technology on a real-world registered timetable in East Scotland. Although passenger adoption did not meet expectations, the trial has significantly advanced the understanding of the operational and regulatory requirements for autonomous services, delivering what was expected from this demonstrator project. The partners remain committed to exploring new opportunities for self-driving technology in other areas across the UK, ensuring that this exciting innovation can play a transformative role in future transport networks."

[Cadavre Exquis]

'Human Vs. Autonomous Car' Race Ends Before It Begins
Posted by EditorDavid on Sunday December 22, 2024 @04:47PM from the thunder-alley dept.

A demonstration "race" between a (human) F1 race car driver Daniil Kvyat and an autonomous vehicle was just staged by the Abu Dhabi Autonomous Racing League.

Describing the league and the "man vs. machine" showdown, Ars Technica writes, "Say goodbye to the human driver and hello to 95 kilograms of computers and a whole suite of sensors."

But again, racing is hard, and replacing humans doesn't change that. The people who run and participate in A2RL are aware of this, and while many organizations have made it a sport of overselling AI, A2RL is up-front about the limitations of the current state of the technology. One example of the technology's current shortcomings: The vehicles can't swerve back and forth to warm up the tires. Giovanni Pau, Team Principal of TII Racing, stated during a press briefing regarding the AI system built for racing, "We don't have human intuition. So basically, that is one of the main challenges to drive this type of car. It's impossible today to do a correct grip estimation. A thing my friend Daniil (Kvyat) can do in a nanosecond...."

Technology Innovation Institute (TII) develops the hardware and software stack for all the vehicles. Hardware-wise, the eight teams receive the same technology. When it comes to software, the teams need to build out their own system on TII's software stack to get the vehicles to navigate the tracks. In April, four teams raced on the track in Abu Dhabi. As we've noted before, how the vehicles navigate the tracks and world around them isn't actually AI. It's programmed responses to an environment; these vehicles are not learning on their own. Frankly, most of what is called "AI" in the real world is also not AI.

Vehicles driven by the systems still need years of research to come close to the effectiveness of a human beyond the wheel. Kvyat has been working with A2RL since the beginning. In that time, the former F1 driver has been helping engineers understand how to bring the vehicle closer to their limit. The speed continues to increase as the development progresses. Initially, the vehicles were three to five minutes slower than Kvyat around a lap; now, they are about eight seconds behind. That's a lifetime in a real human-to-human race, but an impressive amount of development for vehicles with 90 kg of computer hardware crammed into the cockpit of a super formula car. Currently, the vehicles are capable of recreating 90-95 percent of the speed of a human driver, according to Pau. Those capabilities are reduced when a human driver is also on the track, particularly for safety reasons....

The "race" was to be held ahead of the season finale of the Super Formula season... The A2RL vehicle took off approximately 22 seconds ahead of Kvyat, but the race ended before the practice lap was completed. Cameras missed the event, but the A2RL car lost traction and ended up tail-first into a wall...

Khurram Hassan, commercial director of A2RL, told Ars that the cold tires on the cold track caused a loss of traction.

Saludos.

[Cadavre Exquis]

La movilidad del Unitree B2-W en el, probablemente, vídeo de robots más alucinante del año

Por @Alvy — 24 de diciembre de 2024

Unitree Robotics | Unitree B2-W Talent Awakening!

Este vídeo de demostración muestra las habilidades de la nueva actualización del robot industrial B2-W de Unitree, cuyos modelos ya habíamos visto por aquí con anterioridad. Aquí muestra una agilidad y capacidad de movimientos todoterreno asombrosas, que han dejado a todo el mundo con la boca abierta.

Y es que podemos sin duda llamarlo blackmirroriano, inquietante y decir que da miedito, pero también que es el «vídeo de robots del año». Con esa movilidad es difícil hasta para los humanos y los drones voladores seguirlo; nada que decir además cuando se lanza por terraplenes o se deja caer casi dos metros sin problema. Y encima con movimientos elegantes. Casi se diría que de patinador olímpico.

Y ya que alguien se ponga a montarlo a caballito, como quien dice, por mucho que lleve casco de bici… Pues qué decir, que igual parece una temeridad, aunque ellos sabrán lo que se hacen y habrán hecho mil ensayos previos. Mientras lo veía me imaginaba que estos chismes pueden dar lugar a nuevos deportes, carreras de competición… En fin, lo típico cuando chismes alucinantes consiguen comportarse de forma más o menos estable.

El vídeo parece que no tiene trucos (excepto quizá el de «grabo muchos intentos y me quedo con los que mejor resultan). Las escenas tienen sus sombras al amanecer –hora típica para que queden mejor iluminadas– las piedras levantan polvo, también las ruedas… Salen personas en los mismos planos, hay más vídeos anteriores… Nada que objetar. Felicidades al equipo de Unitree por el vídeo y el B2-W y a ver cómo estará este modelo cuando acabe 2025… Igual hasta limpia la casa, recoge la mesa y lava los platos.

Saludos.

[Cadavre Exquis]

New Physics Sim Trains Robots 430,000 Times Faster Than Reality
Posted by BeauHD on Monday December 23, 2024 @10:30PM from the exponential-development dept.

An anonymous reader quotes a report from Ars Technica:

On Thursday, a large group of university and private industry researchers unveiled Genesis, a new open source computer simulation system that lets robots practice tasks in simulated reality 430,000 times faster than in the real world. Researchers also plan to introduce an AI agent to generate 3D physics simulations from text prompts. The accelerated simulation means a neural network for piloting robots can spend the virtual equivalent of decades learning to pick up objects, walk, or manipulate tools during just hours of real computer time.

"One hour of compute time gives a robot 10 years of training experience. That's how Neo was able to learn martial arts in a blink of an eye in the Matrix Dojo," wrote Genesis paper co-author Jim Fan on X, who says he played a "minor part" in the research. Fan has previously worked on several robotics simulation projects for Nvidia. [...] The team also announced they are working on the ability to generate what it calls "4D dynamic worlds" -- perhaps using "4D" because they can simulate a 3D world in motion over time. The system will reportedly use vision-language models (VLMs) to generate complete virtual environments from text descriptions (similar to "prompts" in other AI models), utilizing Genesis's own simulation infrastructure APIs to create the worlds.

Saludos.

[pollo]

New Physics Sim Trains Robots 430,000 Times Faster Than Reality
Posted by BeauHD on Monday December 23, 2024 @10:30PM from the exponential-development dept.

An anonymous reader quotes a report from Ars Technica:

On Thursday, a large group of university and private industry researchers unveiled Genesis, a new open source computer simulation system that lets robots practice tasks in simulated reality 430,000 times faster than in the real world. Researchers also plan to introduce an AI agent to generate 3D physics simulations from text prompts. The accelerated simulation means a neural network for piloting robots can spend the virtual equivalent of decades learning to pick up objects, walk, or manipulate tools during just hours of real computer time.

"One hour of compute time gives a robot 10 years of training experience. That's how Neo was able to learn martial arts in a blink of an eye in the Matrix Dojo," wrote Genesis paper co-author Jim Fan on X, who says he played a "minor part" in the research. Fan has previously worked on several robotics simulation projects for Nvidia. [...] The team also announced they are working on the ability to generate what it calls "4D dynamic worlds" -- perhaps using "4D" because they can simulate a 3D world in motion over time. The system will reportedly use vision-language models (VLMs) to generate complete virtual environments from text descriptions (similar to "prompts" in other AI models), utilizing Genesis's own simulation infrastructure APIs to create the worlds.

Saludos.

Esto es más viejo que el catarro. Como ejemplo, los coches autónomos llevan años "entrenando" con esta idea. Pero como siempre, la realidad es mucho más compleja que cualquier simulación que se quiera hacer, por sofisticada que sea. El problema no es el tiempo de entrenamiento, es todo lo que no se entrena con esto. Y además sigue sin ser suficiente ya que no hay razonamiento real detrás de lo que se "aprende" con lo que una situación imprevista acaba llevando a que el sistema tome decisiones impredecibles o absurdas.

Si fuese por tiempo o fuerza bruta ya tendríamos coches (o robots de cualquier tipo) que se manejasen en todas las situaciones posibles, pero de momento no es así.

[Cadavre Exquis]

https://elpais.com/ciencia/materia-gris/2024-12-23/el-cerebro-controla-el-peso-corporal-y-la-obesidad-regulando-la-absorcion-intestinal-de-grasas.html

Saludos.