Meta AI’s new “BlenderBot”: The next generation!

Meta AI said a cou­ple of days ago that it released a new ver­sion of its advanced Blender­Bot chat­bot that is able to remem­ber pre­vi­ous inter­ac­tions and learn from them.

In a blog post, Meta AI researchers said the upgrad­ed chat­bot can search the inter­net for infor­ma­tion so it can chat about almost any top­ic, while improv­ing its con­ver­sa­tion­al skills through nat­ur­al con­ver­sa­tions and feed­back “in the wild.”

Accord­ing to Meta AI, Blender­Bot 3 is said to be the world’s first 175 bil­lion-para­me­ter, pub­licly avail­able chat­bot that comes com­plete with mod­el weights, code, datasets and mod­el cards.

The orig­i­nal Blender­Bot, which Meta AI launched two years ago, had the abil­i­ty to blend skills such as empa­thy, knowl­edge and per­son­al­i­ty into a com­plete AI sys­tem. One year after that, Meta AI launched Blender­Bot 2, in which researchers had added a long-term mem­o­ry capa­bil­i­ty that enabled it to hold more engag­ing and sophis­ti­cat­ed con­ver­sa­tions on vir­tu­al­ly any topic.

Meta AI’s long-term goal is to build more real­is­tic AI sys­tems that can inter­act with humans in more intel­li­gent, use­ful and safer ways, and to do this it says it must adapt the mod­els that pow­er them to our ever-chang­ing needs.

The unit claims that Blender­Bot 3 deliv­ers supe­ri­or per­for­mance to any chat­bot because it’s based on its pub­licly avail­able OPT-175B lan­guage mod­el, which is 58 times larg­er than the mod­el that pow­ered Blender­Bot 2.


“Most pre­vi­ous pub­licly avail­able datasets are typ­i­cal­ly col­lect­ed through research stud­ies with anno­ta­tors that can’t reflect the diver­si­ty of the real world,” Meta AI explained.

Read: Time Travel Using AI + VR?

Through a live, pub­lic demo, so far only avail­able in the U.S., Blender­Bot 3 can learn from inter­ac­tions with any­one. The expe­ri­ence it gains from these con­ver­sa­tions will enable it to hold longer and more diverse con­ver­sa­tions, Meta AI said, and pro­vide more var­ied feed­back. For instance, those who chat with it can pro­vide feed­back to each response with a thumbs up or down, spec­i­fy­ing what they didn’t like about each neg­a­tive com­ment, such as because it was off-top­ic, rude, spam­like, non­sen­si­cal or some­thing else.

Blender­Bot 3 also takes steps to address the real­i­ty that not every­one who is using it will have good inten­tions. To that end, it incor­po­rates learn­ing algo­rithms aimed at dis­tin­guish­ing between help­ful and harm­ful feedback.

“We hope this work will help the wider AI com­mu­ni­ty spur progress in build­ing ever-improv­ing intel­li­gent AI sys­tems that can inter­act with peo­ple in safe and help­ful ways,” the researchers said.

The next generation of BlenderBot

As the first in the series of Bleder­Bot was just a toy, the sec­ond one was a step for­ward, with long-term mem­o­ry and the vocab­u­lary, which had grown to 200k words. And now, Blender­Bot 3 has long-term mem­o­ry capac­i­ty and even the abil­i­ty of self-learning.

That means that the next gen­er­a­tion of Blender­Bot will pos­si­bly have cog­ni­tive archi­tec­ture: knowl­edge of facts about the envi­ron­ment; being able to learn from inter­ac­tions with peo­ple in real time; mod­els of per­cep­tion and cog­ni­tion based on data from exter­nal sources; per­son­al­i­ty and emo­tions with para­me­ters that can vary depend­ing on circumstances.

In addi­tion, the next gen­er­a­tion of Blender­Bot will be able to inter­act with peo­ple based on its abil­i­ty to gen­er­ate some­thing more than just a pre­set answer, such as demon­strat­ing a sense of humor.

On the oth­er hand, Google’s Brain Team have announced Ima­gen, a text-to-image AI mod­el that can gen­er­ate pho­to­re­al­is­tic images of a scene giv­en a tex­tu­al descrip­tion and DALL·E 2, which is a new AI sys­tem that can cre­ate real­is­tic images and art from a descrip­tion in nat­ur­al language.

Source here

Leave a Reply

Your email address will not be published.

Join our NewsletterDaily Glimple of Future

Our blog, "Daily Glimpse of Future", strives to make the future much clearer than it is today. Join our newsletter for free now.