Boston Dynamics updates Spot robot dog; Meta unveils a GPT-3 rival, and other AI news

We aim to inform readers not only about developments in the bitcoin industry, but also about what is happening in adjacent technological fields — cybersecurity and the world of artificial intelligence (AI).

ForkLog AI has gathered the most important AI news from the past week.

Google allows users to remove personal information from search results

Google has provided users the ability удалить personal information from search results.

Under the updated privacy policy, upon request, the system will stop displaying certain pages containing the user’s personal data. Google may remove:

• confidential government identification data, including the Social Security number;
• bank account numbers;
• credit card numbers;
• images of handwritten signatures;
• photos of identity documents;
• medical records;
• personal contact information, including address, phone number and email;
• data to access any account.

Google also promised to delete intimate personal photos, pornographic deepfakes and images depicting violence.

Meta unveils open-source language model

Meta опубликовала code for the 175-billion-parameter OPT-175B language model. The neural network доступна to AI researchers on request for non-commercial use.

In a statement, the company said that they share a model trained on public data to help the community understand the principles of this technology. Meta did not rule out that scientists could remove toxicity and bias from models.

Researchers trained the OPT-175B neural network using 992 Nvidia 80GB A100 GPUs, achieving 147 TFLOPS on the chip. According to Meta, researchers do not need to build an algorithm and train it from scratch, as the tech giant provides code to deploy it on 16 Nvidia V100 GPUs.

The company also открыла access to smaller pre-trained models up to 66 billion parameters for all comers.

Volkswagen signs partnership with Qualcomm on autonomous driving

Autokoncern Volkswagen signed a multiyear collaboration with chipmaker Qualcomm to develop autonomous driving technology.

Details of the deal with Qualcomm are not disclosed. It is known that the company will supply system-on-a-chip (SoC) for the Cariad platform, enabling fourth-level autonomous driving. Volkswagen will integrate the processors into the onboard computers of all brands within the group.

The carmaker also said it is in talks with Israeli supplier Mobileye.

PrivatBank launches voice biometrics

The Ukrainian bank PrivatBank launched a voice authentication system for customers calling the hotline.

To use the feature, a user must create a voice print via the Privat24 app. After that, on subsequent calls to the hotline the algorithm can quickly recognise the customer.

After enabling voice biometrics, users will not have to answer additional security questions, the bank said.

YouTube Kids algorithms push content about drugs and weapons

The Tech Transparency Project found that the YouTube Kids algorithms recommend videos to children about weight loss, skin whitening, drug subcultures and firearms.

Researchers created three accounts for one age group. Among the videos for “younger children” was a cooking show themed around “Breaking Bad” and a process of building a motorhome in Minecraft, “where crystal meth is being cooked”.

The platform also showed younger children videos about recoil pads that protect shooters from weapon recoil, and for older children a guide to building a shelf with a hidden compartment for a pistol.

The company said that thanks to automated filters, human moderation and parental feedback, YouTube Kids videos are suitable for family viewing, but “no system is perfect.” After review, developers removed or age-restricted the flagged videos in the Kids app.

Google fires AI researcher over criticism of the company

Google fired another AI researcher amid criticism of the tech giant’s work.

According to The New York Times, in March 2022 the company sidelined researcher Satrajit Chatterjee for challenging a paper arguing that algorithms could design chips more efficiently than people.

Company representatives declined to comment on the dismissal, but noted it happened “for a legitimate reason”. The original paper had been “carefully checked” and peer-reviewed, and the work challenging it “did not meet the company’s standards,” they added.

Earlier Google sidelined two leading AI ethics researchers Timnit Gebru and Margaret Mitchell, which led to a series of other voluntary exits.

The robocar sets a new speed record for autonomous vehicles

The PoliMOVE team’s robocar shattered the world speed record for autonomous vehicles, reaching 309.3 km/h.

The vehicle ran on a straight takeoff-and-landing strip at NASA’s Kennedy Space Center. The robocar reached 309.3 km/h over 1 km in two runs in opposite directions to mitigate wind effects.

Meta uses AI to design “green” concrete

Meta will build new data-centers for the metaverse out of “green” concrete, created with AI. The material formulation trimmed carbon emissions by 40% and exceeded strength requirements.

The tech giant, in collaboration with researchers from the University of Illinois Urbana-Champaign, MIT and IBM, trained the model on a dataset containing 1,000 concrete formulas and structural characteristics, including 7- and 28-day compressive strength data. The team also measured the carbon footprint of the resulting mixes.

They then selected five of the most promising variants and iteratively refined them. In the end, they derived a formula replacing 50% of cement with fly ash and slag.

Meta partnered with Ozinga to optimise the mix and test it under real-world conditions.

Boston Dynamics updates Spot robot dog

Boston Dynamics showcased the Spot in a factory in a new video “No Time To Dance.”

In the clip, the four-legged machine carries out a series of tasks in industrial settings. The robot takes sensor readings, scans equipment with an infrared camera, generates work orders for staff and even dances.

Boston Dynamics also showcased updates to the robot dog.

Among the improvements:

• five stereo cameras capturing full-colour images and providing enhanced scene-depth data;
• a tablet controller with a larger screen and longer battery life;
• a new system-on-chip for accelerating on-device AI processing and 5G compatibility.

Researchers develop algorithm to identify actions in video

MIT engineers developed an AI algorithm to identify actions in video and label them without human input.

The model learns to represent data by capturing concepts common to visual and auditory modalities. The system will determine that a crying child in a video is associated with the spoken word “cry” in the audio clip.

Researchers used attribute learning. The algorithm receives raw data, including video and corresponding text captions, and encodes it, extracting features of objects and actions from the clip.

Then the AI maps the data into a representation space. The model clusters similar information together as discrete points, expressing each with a single word. The researchers limited the model to 1,000 words for labeling vectors.

The approach also uses a shared representation space. This allows the algorithm to study the relationship between a juggling video and the audio track containing the word “juggles”.

The week’s top AI deals

From May 1 to 7, 2022, AI startups raised over $720m. Here are the most notable deals.

• Anthropic собрала $580m for research aimed at making AI more reliable and explainable.
• Traceable AI получила $60m to provide services for protecting APIs from cyberattacks.
• Concerto, a developer of credit cards for brands, привлек $21.2m.
• Accern получила $20m to develop AI apps for analyzing online financial documents.
• SLAMcore собрали $16m for spatial understanding in real-world robots and the metaverse.
• Druid привлекла $15m to scale a platform for building chatbots.
• Kinetix получили $11m.
• Galileo собрала $5.1m to advance a service for streamlining AI-model development.
• RECON Labs привлекла $4.4m to enhance a platform turning 2D product images into 3D models.

Also on ForkLog:

• DeepMind presented visual language model with 80 billion parameters.
• AI invented an enzyme to break down plastic in 24 hours.
• Copilot “leaked” private keys from crypto wallets.
• Testing has started in the US for the first commercial neurointerface.
• Apple, Google and Microsoft will implement a passwordless sign-in standard in online services.

Subscribe to ForkLog AI on TikTok!