Parlons Futur
Exemple concret de comment l’IA ne signifie pas forcément la fin du libre-arbitre
OpenAI et la main robot qui résout le Rubik’s cube : décryptage derrière la hype

OpenAI et la main robot qui résout le Rubik’s cube : décryptage derrière la hype

October 25, 2019

Résumé ci-dessous issu de cet épisode de ma newsletter ParlonsFutur

  • Il faut noter que la main robotique elle-même en tant que hardware existe depuis 15 ans, utilisée par de nombreux labos pour leurs expériences

  • l'algorithme lui-même qui permet de savoir quelles faces tourner dans quel sens et dans quel l'ordre, bref l'algo qui permet de résoudre le Rubik's cube, existe depuis 17 ans, un algo codé à la main, et non découvert par apprentissage par la machine, ce n'est pas ce dont il est question ici.

  • De nombreux robots parviennent à résoudre le Rubik's cube, plus vite même que celui d'OpenAI ici (par ex, “a machine developed by MIT solved a cube in less than 0.4 seconds”)

  • Mais aucun robot main humanoïde de ce type y était parvenu (pour un robot avec 24 degrés de liberté dans les mouvement, contre 7 en général pour un bras robot de base), c'est la dextérité du robot main qui est ici révolutionnaire, et la façon dont l'apprentissage s'est fait

  • ici la nouveauté est que le logiciel derrière la main a appris de lui-même comment le faire, via un "apprentissage par renforcement", c'est-à-dire en sachant quel est l'objectif à atteindre, et en se faisant récompenser ou pénaliser plus il se rapproche ou s'éloigne de l'objectif

  • “Human hands let us solve a wide variety of tasks. For the past 60 years of robotics, hard tasks which humans accomplish with their fixed pair of hands have required designing a custom robot for each task. As an alternative, people have spent many decades trying to use general-purpose robotic hardware, but with limited success due to their high degrees of freedom. In particular, the hardware we use here is not new—the robot hand we use has been around for the last 15 years—but the software approach is.”

  • l'autre révolution est que le robot n'a pas eu besoin de tout apprendre dans le monde réel, mais dans une simulation virtuelle, ce qui accélère l'apprentissage puisqu'on peut accélerer le temps dans les simulations et en tenir de très nombreuses en parallèle !

    • OpenAI put it through thousands of years of training in simulation before trying the physical cube solve.

    • But this only takes a few days because we can parallelize the training.

    • You also don’t have to worry about the robots breaking or hurting someone as you’re training these algorithms,”

  • "we developed a new method called Automatic Domain Randomization (ADR), which endlessly generates progressively more difficult environments in simulation. This frees us from having an accurate model of the real world, and enables the transfer of neural networks learned in simulation to be applied to the real world”

  • ADR starts with one environment, wherein a neural network learns to solve Rubik’s Cube. As the neural network gets better at the task and reaches a performance threshold, the amount of domain randomization is increased automatically. This makes the task harder, since the neural network must now learn to generalize to more randomized environments.

  • One of the parameters we randomize is the size of the Rubik’s Cube. ADR begins with a fixed size of the Rubik’s Cube and gradually increases the randomization range as training progresses. We apply the same technique to all other parameters, such as the mass of the cube, the friction of the robot fingers, and the visual surface materials of the hand. The neural network thus has to learn to solve the Rubik’s Cube under all of those increasingly more difficult conditions. ADR exposes the network to an endless variety of randomized simulations

  • not only did they change how much gravity there is in the simulation—they changed which way gravity points.

  • We find that our system trained with ADR is surprisingly robust to perturbations even though we never trained with them: The robot can successfully perform most flips and face rotations under all tested perturbations, though not at peak performance.

  • The system can handle situations it never saw during training (comme être embêté par une girafe en peluche, voir vidéo, ou avoir des doigts attachés, etc.) This shows that reinforcement learning isn’t just a tool for virtual tasks, but can solve physical-world problems requiring unprecedented dexterity.

  • Our robot still hasn’t perfected its technique though, as it solves the Rubik’s Cube 60% of the time (and only 20% of the time for a maximally difficult scramble).

  • Un pas vers une IA (et des robots) capable de généraliser ??

    • We believe that meta-learning, or learning to learn, is an important prerequisite for building general-purpose systems, since it enables them to quickly adapt to changing conditions in their environments. The hypothesis behind our methodology ADR is that a memory-augmented networks combined with a sufficiently randomized environment leads to emergent meta-learning, where the network implements a learning algorithm that allows itself to rapidly adapt its behavior to the environment it is deployed in.

    • "We believe that human-level dexterity is on the path towards building general-purpose robots"

  • Un pas vers l'explicabilité ? (et la fin du syndrome boîte noire absolue?)

    •  Ils ont développé un moyen de visualiser ce qu'il se passe dans la mémoire de l'algo, en temps réel

    •  "We find that each memory group has a semantically meaningful behavior associated with it. For example, we can tell by looking at only the dominant group of the network’s memory if it is about to spin the cube or rotate the top clockwise before it happens."

    • Visualizing our networks enables us to understand what they are storing in memory.

    • For example, we can tell by looking at only the dominant group of the network’s memory if it is about to spin the cube or rotate the top clockwise before it happens

  • Voir une vidéo du robot résolvant un cube, non éditée, vitesse réelle!

  • Voir la vidéo de présentation des travaux avec toutes les perturbations originales surmontées et non-rencontrées au préalable pendant l'apprentissage

  • Voir l'article sur le site d'OpenAI, avec plein de vidéos sympas

Interview of Nicolas Colin, co-founder of The Family and author of Hedge: A Greater Safety Net for the Entrepreneurial Age

Interview of Nicolas Colin, co-founder of The Family and author of Hedge: A Greater Safety Net for the Entrepreneurial Age

September 19, 2019
Nicolas Colin is the co-founder of The Family, a company that helps European startups scale. He's also the author of Hedge: A Greater Safety Net for the Entrepreneurial Age
 
[full transcript of the Q&A can be found here]
 
Q&A between Nicolas Colin (twitter.com/Nicolas_Colin) & Thomas Jestin (twitter.com/thomasjestin) in Singapore, March 2019
  • 00:00 : intro : who's Nicolas Colin, his background, why he's in Singapore for a few days
  • 2min 40sec :  tell us what The Family is
  • 7m 35s : can you comment on some of the points of The Family's manifesto ?
  • 8m 28s : how The Family was born from the idea that we're now in the Entrepreneurial Age (and not so much the information age), derived from that foundational article The Entrepreneurial age by Babak Nivi, cofounder of AngelList
  • 11m 00s : what is an entrepreneur ?
  • 13m 50s : The Family has the ambition to become the new Berkshire Hathaway (as detailed here), by providing capital and unfair advantages to entrepreneurs, what about these unfair advantages ? (less and less about these, more about capital and relationships with VCs, access to VCs, help raise fast and on good terms, they have a white list of VCs)
  • 16m 50s : explain why you don't believe in gathering startups together under the same roof, like in the case of France at Station F
  • 20m 52s : why don't you believe in mentors for European startups ?
  • 24m 30s : can you say a word about some of the startups you're helping that are representative of your portfolio ? (special mention about Heetch, a French VTC app, see more here)
  • 29m 00s : how Nicolas Colin was sued himself by the CEO of the leading taxi company in Paris
  • 30m 50s : tell us a bit about how you used growth hacking yourself at The Family
  • 36m 38s : about ramen-entrepreneurs and cockroach-entrepreneurs
  • 38m 15s : quick intro of the book Hedge: A Greater Safety Net for the Entrepreneurial Age
  • 39m 40s : about the Tech Backlash that started in the US a few years ago, as the genesis for the book
  • 42m 55s : the argument of the book : we have to reinvent the Great Safety Net
  • 44m 05s : what the first Great Safety Net was about (see the illustration on that article with full transcript of the Q&A)
  • 46m 45s : why the Great Safety Net doesn't work anymore
  • 48m 42s : please define your concept of "the multitude" (or the "networked individual"), different in your mind from the masses and the people
  • 53m 33s : why our age is not so much the age of data, but the Entrepreneurial Age
  • 55m 52s : how now power lies outside of companies, not inside anymore
  • 56m 55s : tell us about your definition of a tech company (clue : telco are not tech companies) (see image in the article with transcript of the Q&A)
  • 1h 00m 07s : what about that Great Safety Net 2.0 (see image in the article)
  • 1h 02m 35s : tell us about your concept of hunters vs settlers (see image in the article)
  • 1h 09m 40s : the particular case of Singapore that wants to attract hunters but at the same time prevent its own people from becoming hunters themselves and leaving Singapore
  • 1h 11m 55s : please elaborate on that idea of "exit unions", as a cog in your Great Safety Net 2.0
  • 1h 16m 58s : why entrepreneurs will be the ones bringing about that Greater Safety Net, and not the state : the example of Lambda school
  • 1h 18m 11s : examples of start-ups from The Family's portfolio helping bring about the Greater Safety Net (www.side.co)
  • 1h 19m 20s : why do you think the Universal Basic Income is such a bad idea ?
  • 1h 22m 56s : about BREXIT, Boris Johnson said that the UK can become the new Singapore, is it going to happen ? is BREXIT affecting your work ? (mention of How Asia Works, by Joe Studwell)
  • 1h 25m 34s : why GDPR doesn't work
Questions from the public
  • 1h 28m 20s : apart from market size and homogeneity and the fact that European governments don't understand startups, why don't we have European Googles or Facebooks ? (clue : we have too good a life in Europe)
  • 1h 32m 48s : are Alibaba and Tencent tech companies based on your definition ?
  • 1h 34m 26s : you talked a lot about consumer-based businesses, what about the future of enterprise businesses (businesses targeting businesses) in your opinion ?
  • 1h 37m 12s : do you think that the networked individual system works or already works in China ? does it have potential in Africa in the future as well ?
  • 1h 39m 55s : what are your expectations from the ideal investor ? (in short : swift response, good terms, support when in trouble)
  • 1h 42m 01s : what are you looking for when assessing an early-stage startup to help and invest in ?  (in short : intensity)

[full transcript of the Q&A can be found here]

Interview happened with the help of the French Tech Singapore and the ENGIE factory in March 2019

Entretien exclusif avec Jacques Attali : intelligence artificielle, avenir du travail, GAFA, Singularité, dictature volontaire et plus encore

Entretien exclusif avec Jacques Attali : intelligence artificielle, avenir du travail, GAFA, Singularité, dictature volontaire et plus encore

November 12, 2018

00 min 56 sec : présentation du parcours de Jacques Attalli (JA)

01 min 56 sec : présentation de son oeuvre écrite

03 min 55 sec : comment JA définit l'intelligence artificielle (IA)

04 min 11 sec : comment définissez-vous l'intelligence humaine ?

09 min 36 sec : faut-il découpler intelligence et conscience ? peut-on imaginer des machines intelligentes qui ne seraient pas conscientes ?

11 min 35 sec :pourquoi l'IA n'est qu'une "technique sommaire et utile de prédiction, étape assez passagère avant l'avènement de sciences et technologies infiniment plus prometteuses", et constitue en fait "un appauvrissement majeur de la réflexion intellectuelle" selon JA

14 min 28 sec : l'IA va-t-elle renforcer l'emprise des GAFA sur le monde ?

15 min 05 sec : "l'IA jouera un rôle encore plus importants dans les pays totalitaires voire dans la mise en place de systèmes totalitaires dans les pays démocratiques car.."

16 min 02 sec : "l'IA sera d'abord une technologie chinoise parce que..."

18 min 06 sec : L'IA induira-t-elle un chômage de masse ?

22 min 01 sec :  Sachant que l'IA sera bientôt en mesure de comprendre nos émotions et de les simuler à la perfection, les métiers reposant sur le rapport humain sont-ils donc vraiment à l'abri ?

23 min 55 sec : aux Etats-Unis, de plus en plus de jeunes choisissent délibérément de se retirer du marché du travail pour mieux profiter d'une offre de loisirs numériques toujours plus vaste, qualititative et meilleure marché voire gratuite, que pensez-vous de cette tendance ?

27 min 12 sec : que faire face aux nouvelles addictions comme les jeux vidéos et les réseaux sociaux ?

30 min 22 sec : croyez-vous à la théorie de la Singularité ? "Monsieur Kurzweil m'a toujours paru comme un gentil naïf manquant infiniment de culture"

Pourquoi le scénario dystopien de Harari est improbable!
Faut-il interdire les armes autonomes? Et en particulier les micro-drones autonomes et tueurs
Quelques définitions: IA, Machine Learning, Deep Learning, Supervised and Unsuopervised Learning, Reinforcement Learning
Éléphants, miroirs des Hommes

Éléphants, miroirs des Hommes

December 25, 2017
 
Meilleurs extraits de l'article de The Economist Conserve elephants. They hold a scientific mirror up to humans

 
 

Elephants, about as unrelated to human beings as any mammal can be, seem nevertheless to have evolved intelligence, and possibly even consciousness. Though they may not be alone in this (similar claims are made for certain whales, social carnivores and a few birds), they are certainly part of a small and select group. Losing even one example of how intelligence comes about and makes its living in the wild would not only be a shame in its own right, it would also diminish the ability of biologists of the future to understand the process, and thus how it happened to human beings.

 
Dr Wittemyer argues that, human beings aside, no species on Earth has a more complex society than that of elephants. And elephant society does indeed have parallels with the way humans lived before the invention of agriculture.
 
The nuclei of their social arrangements are groups of four or five females and their young that are led by a matriarch who is mother, grandmother, great-grandmother, sister or aunt to most of them. Though males depart their natal group when maturity beckons at the age of 12, females usually remain in it throughout their lives.
 
Families are part of wider “kinship” groups that come together and separate as the fancy takes them. Families commune with each other in this way about 10% of the time. On top of this, each kinship group is part of what Dr Douglas-Hamilton, a Scot, calls a clan. Clans tend to gather in the dry season, when the amount of habitat capable of supporting elephants is restricted. Within a clan, relations are generally friendly. All clan members are known to one another and, since a clan will usually have at least 100 adult members, and may have twice that, this means an adult (an adult female, at least) can recognise and have meaningful social relations with that many other individuals.
 
A figure of between 100 and 200 acquaintances is similar to the number of people with whom a human being can maintain a meaningful social relationship—a value known as Dunbar’s number, after Robin Dunbar, the psychologist who proposed it. Dunbar’s number for people is about 150. It is probably no coincidence that this reflects the maximum size of the human clans of those who make their living by hunting and gathering, and who spend most of their lives in smaller groups of relatives, separated from other clan members, scouring the landscape for food.
 
Dealing with so many peers, and remembering details of such large ranges, means elephants require enormous memories. Details of how their brains work are, beyond matters of basic anatomy, rather sketchy. But one thing which is known is that they have big hippocampuses. These structures, one in each cerebral hemisphere, are involved in the formation of long-term memories. Compared with the size of its brain, an elephant’s hippocampuses are about 40% larger than those of a human being, suggesting that the old proverb about an elephant never forgetting may have a grain of truth in it.
 
In the field, the value of the memories thus stored increases with age. Matriarchs, usually the oldest elephant in a family group, know a lot. The studies in Amboseli and Samburu have shown that, in times of trouble such as a local drought, this knowledge permits them to lead their groups to other, richer pastures visited in the past. Though not actively taught (at least, as far as is known) such geographical information is passed down the generations by experience. Indeed, elephant biologists believe the ability of the young to benefit by and learn from the wisdom of the old is one of the most important reasons for the existence of groups—another thing elephants share with people.
 
Nor is it only in their social arrangements that elephants show signs of parallel evolution with humans. They also seem to have a capacity for solving problems by thinking about them in abstract terms. This is hard to demonstrate in the wild, for any evidence is necessarily anecdotal. But experiments conducted on domesticated Asian elephants (easier to deal with than African ones) show that they can use novel objects as tools to obtain out-of-reach food without trial and error beforehand. This is a trick some other species, such as great apes, can manage, but which most animals find impossible.
 
Wild elephants engage in one type of behaviour in particular that leaves many observers unable to resist drawing human parallels. This is their reaction to their dead. Elephant corpses are centres of attraction for living elephants. They will visit them repeatedly, sniffing them with their trunks and rumbling as they do so (see picture overleaf). This is a species-specific response; elephants show no interest in the dead of any other type of animal. And they also react to elephant bones, as well as bodies, as Dr Wittemyer has demonstrated. Prompted by the anecdotes of others, and his own observations that an elephant faced with such bones will often respond by scattering them, he laid out fields of bones in the bush. Wild elephants, he found, can distinguish their conspecifics’ skeletal remains from those of other species. And they do, indeed, pick them up and fling them into the bush.
 
 
Futur de la reproduction: utérus artificiel, pouvoir être à la fois le père et la mère génetique d’un enfant
Quand un singe arrive à contrôler un 3ème bras par la pensée