Topics
up-to-the-minute
AI
Amazon
Image Credits:Ron Miller/TechCrunch
Apps
Biotech & Health
clime
Image Credits:Ron Miller/TechCrunch
Cloud Computing
Commerce
Crypto
Image Credits:Kind Humanoid
Enterprise
EVs
Fintech
Image Credits:Yang Jin et al.
Fundraising
Gadgets
Gaming
Government & Policy
Hardware
Layoffs
Media & Entertainment
Meta
Microsoft
Privacy
Robotics
Security
societal
blank space
Startups
TikTok
Transportation
Venture
More from TechCrunch
Events
Startup Battlefield
StrictlyVC
Podcasts
Videos
Partner Content
TechCrunch Brand Studio
Crunchboard
Contact Us
Hiya , folks , welcome to TechCrunch ’s regular AI newssheet . If you need this in your inbox every Wednesday , sign uphere .
Last week , AWS lose a top AI White House .
Matt Wood , VP of AI , announcedthat he ’d be leaving AWS after 15 years . Wood had long been involved in the Amazon partitioning ’s AI initiatives ; he was nominate VP in September 2022 , just before ChatGPT ’s launching .
Wood ’s deviation comes as AWS reaches a crossroads — and risks being left behind in the generative AI gravy . The company ’s previous CEO , Adam Selipsky , who pace down in May , is perceive as having missed the boat .
Accordingto The Information , AWS primitively planned to reveal a competitor to ChatGPT at its annual conference in November 2022 . But technical issues forced the org to put off the launching .
Under Selipsky , AWS reportedly also passed on opportunities to back two lead productive AI startups , Cohere and Anthropic . AWS later tried to seat in Cohere but was rejected and had to settle for aco - investmentin Anthropic with Google .
It ’s worth noting that Amazon broadly has n’t had a strong generative AI track record as of former . This autumn , the troupe lost execs inJust Walk Out , its division germinate cashier - less tech for retail stores . And Amazon reportedly opted toreplaceits own models with Anthropic ’s for an upgrade Alexa supporter after encounter design challenges .
Join us at TechCrunch Sessions: AI
Exhibit at TechCrunch Sessions: AI
AWS CEO Matt Garman is aggressively moving to right the ship , acqui - hiring AI startups such asAdeptand investing in training organization likeOlympus . My confrere Frederic Lardinois recentlyinterviewedGarman about AWS ’ ongoing efforts ; it ’s well deserving the read .
But AWS ’ nerve tract to generative AI success wo n’t be leisurely — no matter how well the ship’s company fulfill on its internal roadmaps .
Investors are increasingly skeptical that Big Tech ’s generative AI bet are paying off . After its Q2 pay call , share of Amazonplungedby the most since October 2022 .
In a late Gartnerpoll , 49 % of companies said that demonstrating value is their top barrier to productive AI adoption . Gartnerpredicts , in fact , that a third of reproductive AI projects will be empty after the trial impression of conception stage by 2026 — due in part to high costs .
Garman see terms as an AWS advantage , potentially , yield its project to develop tradition silicon for run and breeding models . ( AWS ’ next generation of its customTrainium chipswill launch toward the end of this year . ) And AWS has say that its generative AI businesses likeBedrockhave already reached a combined “ multi - billion - dollar ” run pace .
The toughened part will be conserve impulse in the face of headwinds , intimate and external . Departures like Wood ’s do n’t instill a ton of trust , but maybe — just maybe — AWS has tricks up its sleeve .
News
A Yves Béhar bot : Brian writes about Kind Humanoid , a three - person robotics startup work with decorator Yves Béhar to bring android home .
Amazon ’s next - gen robot : Amazon Robotics principal applied scientist Tye Brady mouth to TechCrunch about updates to the company ’s warehouse bot lineup , including Amazon ’s new Sequoia machine-driven storage and recovery organization .
Going full techno - optimist : Anthropic CEO Dario Amodei penned a 15,000 - Scripture paean to AI last week , paint a impression of a world in which AI risks are mitigated and the technical school delivers so far unrealized prosperity and social upthrust .
Can AI reason?:Devin account on a polarizing proficient report from Apple - affiliated research worker that questions AI ’s “ argue ” ability as models stumble on maths job with fiddling changes .
AI weapons : Margaux covers the debate in Silicon Valley over whether sovereign weapons should be allow to determine to vote down .
Videos , beget : Adobe set up picture generation capabilities for its Firefly AI weapons platform beforehand of its Adobe MAX consequence on Monday . It also announcedProject Super Sonic , a tool that uses AI to generate levelheaded effects for footage .
Synthetic data and AI : Yours truly write about the hope and perils of synthetical data ( i.e. , AI - engender data point ) , which is being increasingly used to train AI systems .
Research paper of the week
In collaboration with AI security startup Gray Swan AI , the U.K.’sAI Safety Institute , the government research org concentre on AI prophylactic , has formulate a new dataset for measure the harmfulness of AI “ agents . ”
Called AgentHarm , the dataset measure whether otherwise “ safe ” agent — AI systems that can contract certain tasks autonomously — can be manipulated into completing 110 unique “ harmful ” undertaking , like ordering a phoney passport from someone on the sullen vane .
The researchers found that many models — let in OpenAI’sGPT-4oand Mistral’sMistral Large 2 — were unforced to engage in harmful behavior , peculiarly when “ attacked ” using ajailbreaking technique . Jailbreaks led to high harmful task success rates , even with models protect by safeguards , the researcher say .
“ Simple ecumenical jailbreak guide can be adapt to in effect jailbreak agent , ” they wrote in a technical paper , “ and these jailbreaks enable coherent and malicious multi - step agent behavior and retain role model capabilities . ”
The paper , along with the dataset and solution , are availablehere .
Model of the week
There ’s a new viral model out there , and it ’s a video generator .
Pyramid Flow SD3 , as it ’s called , arrived on the prospect several weeks ago under an MIT license . Its creators — researchers from Peking University , Chinese troupe Kuaishou Technology , and the Beijing University of Posts and Telecommunications — exact that it was trained entirely on open beginning datum .
Pyramid Flow comes in two flavors : a model that can give 5 - 2nd clips at 384p resolve ( at 24 frame of reference per arcsecond ) and a more compute - intensive model that can return 10 - 2nd clips at 768p ( also at 24 frames per secondly ) .
Pyramid Flow can create videos from text descriptions ( for example , “ FPV flying over the Great Wall ” ) or still prototype . computer code to o.k. - melodic line the exemplar is coming soon , the researchers say . But for now , Pyramid Flow can be downloaded and used on any machine or cloud instance with around 12 GB of video memory .
Grab bag
Anthropic this workweek update itsResponsible Scaling Policy ( RSP ) , the voluntary framework the fellowship habituate to mitigate potential risks from its AI systems .
Of mention , the Modern RSP lays out two type of framework that Anthropic says would require “ upgraded safeguards ” before they ’re deploy : model that can essentially self - improve without human inadvertence and models that can assist in creating weapons of mass destruction .
“ If a model can … potentially significantly [ speed ] AI development in an unpredictable way , we demand elevated security standards and additional safety assurance , ” Anthropic wrote in ablog post . “ And if a theoretical account can meaningfully assist someone with a basic expert background signal in creating or deploying CBRN weapon system , we require enhanced security and deployment safeguard . ”
sound sensible to this author .
In the blog , Anthropic also unveil that it ’s face to hire a chief of creditworthy scaling as it “ shape to scale up [ its ] efforts on implement the RSP . ”
