$1 million bet chatGPT won’t lead to AGI, Apple’s smart use of AI, AI millionaires on the rise: AI Eye

$1 Million Bet Chatgpt Won'T Lead To Agi, Apple'S Smart Use Of Ai, Ai Millionaires On The Rise: Ai Eye


Voted By Amazon Polly.

Apple's careful but smart use of AI

Discussion GptDiscussion Gpt
Are you interested in ChatGPT? (Apple)

The big announcement at Apple's WWDC this week might seem unremarkable at first glance, but it's a smart move the more you look at it: Yes, we're finally getting a calculator on the iPad, ten years after the device's debut.

The company uses most of the I.I. It's smartly avoiding issues with products that create problematic illusions, privacy, and promise and give many away.

Instead, Apple is using compact, on-device, open-source models, intelligently hot-swapping adapters to specialize in a specific function, summarization, debugging, or auto-reactions when needed.

bybit

The idea is to increase the likelihood that most AI tasks can be completed successfully and independently on the device – and of course, to finally provide a compelling reason to upgrade your phone or iPad.

More difficult queries will send anonymous, encrypted data to a medium-sized model on Apple's servers (doesn't store the data), and more complex tasks involving typing or artificial intelligence will be sent to ChatGPT after authorization. . OpenAI cannot store your data either.

Based on the specs, the privacy and functionality aspects seem to be heavily considered, though we won't know for sure until September. Apple didn't build the first smartphone, but it brought one of the best versions with the iPhone. It will be interesting to see if his cautiously optimistic approach to AI enjoys similar success.

Apple IntelligenceApple Intelligence
Apple may do some things well, but it's not promising the earth. (Apple)

Google AI is still stuck on glue

Like a snake eating its own tail, all these articles talking about how stupid Google's AI overview answers seem to be making the answer worse. Google searches will remind users that they can eat rocks, that cockroaches live in chickens, and most importantly, that glue is a great way to stick cheese on pizza.

This week, Google AI is still telling people to add two tablespoons of glue to pizza, citing reports from Business Insider and The Verge — about its own wrong answers — as sources. The Verge reporter Elizabeth Lopato wrote:

“Surprising stuff here, folks. Every time someone like me reports on Google AI that something is going wrong, we are training the AI ​​to make mistakes.

Google GlueGoogle Glue
Screenshot of Google Glue Tips. (The Verge)

Two OpenAI researchers predict AGI in 3 years

OpenAI researcher Leopold Aschenbrenner, who was fired for revealing how unprepared the company was for artificial general intelligence, dropped a 165-page document on the matter. It is predicted that AI models will reach the capabilities of human AI researchers and engineers (that is, AGI) by 2027, which could lead to superintelligence when AGI develops its own technology. The prediction is based on the linear progress we have seen in AI in recent years, although critics say the technique may hit a ceiling at some point.

Another OPNI research engineer, James Becker, wrote the same thing: “We basically solved building world models, we're simulating 2-3 years and 1-2 years on System 2 thinking.” The latter two can be done at the same time. Three to five years is estimated for something that “generally looks like an intelligent, embodied agent.”

A French scientist made a $1 million bet that LLMs like ChatGPT do not lead to AGI

AI experts, including META's chief AI scientist Tan Lekun and ASI founder Ben Goertzel, are skeptical that LLMs can provide any kind of path to AGI. On the Dwarkesh Patel podcast this week, French AI researcher Francois Cholet argued that the Open API has set back 5 to 10 years of progress in AGI because frontier research has stopped publishing and the focus on LLM has taken all the oxygen out of the room.

Echoing Lake's highway metaphor, Chollet believes LLMs are “off the road to AGI” and has launched a $1 million ARC prize for any AI system that can pass a four-year Abstraction and Reasoning Corpus (ARC) test. To see if it can adapt to new ideas and situations rather than simply mixing content from the web. Cholet believes that most existing standards simply test memory, which LMMs excel at, and not the ability to creatively grapple with new ideas and situations. It's an interesting philosophical debate: after all, as Patel insists, don't people mostly remember things verbally and generalize or extrapolate from them?

The $1 million prize echoes the skeptic James Randi's $1 million paranormal challenge to anyone who can demonstrate paranormal activity. It was never claimed, and its main purpose was to emphasize that paranormal claims are nonsense. Cholett's intention, however, seems to be to try and focus on general parameters of information rather than memory. Every task on the test is solvable by humans, but not yet by AI.

LLMs cannot reason about novel things

A new study (right) supports the idea that LLMs are surprisingly stupid when it comes to questions people haven't written extensively on the web.

Llm ResearchLlm Research
Alice In Wonderland LLM Research (Arvix)

Despite passing bar tests and other party tricks, the paper found that current LMLMs lack basic reasoning skills and that current metrics fail to accurately detect these deficiencies.

LLMs were asked about this problem: “Alice has N brothers, and she has N sisters. How many sisters does Alice's brother have?”

He concluded: “Using human-solvable common sense (the correct answer is M+1), most of the tested LLMs, including GPT-3.5/4, Claude, Gemini, LLAMA, Mistral and others, show severe failure.” In practice, he often gives nonsensical answers and reasons.

LLMs were also very confident in their wrong answers and provided detailed explanations to justify them.

LeCun highlighted his research in X, “Another possibility is that reasoning and reasoning skills should not be confused with the ability to store and retrieve multiple facts.

LLMs get a lot wrong about the choice.

A study from data analytics startup GroundTruthAI found that Google Gemini 1.0 Pro and various flavors of ChatGPT (from 3.5 to 4o) provide wrong information about voting and the 2024 US election more than a quarter of the time.

The researchers asked 216 multiple-choice questions and found that Gemini 1.0 Pro answered only 57% correctly, while the best OpenAI model (GPT-4o) answered 81% correctly.

The models consistently got Biden and Trump's ages wrong and failed to tell how many days until the election. Two models incorrectly say voters can register on Election Day in Pennsylvania.

Google says its researchers should use an API instead of a public interface, and Wired reports that Google and Microsoft's chatbots are now refusing to answer polling questions.

Also read

Main characteristics

The Invisible Man of the World: How Blockchain Can Bring New Hope to the Stateless Rohingya

Art week

Defying Obsolescence: How Blockchain Tech Can Reshape Artistic Expression

AI training information will end soon

A peer-reviewed study from Epoch AI suggests that technology companies will exhaust the supply of publicly available text-based AI training data between 2026 and 2032.

“There is a serious bottleneck here,” said study co-author Tamai Besiroglu. “If you start hitting those limits of how much data you have, you can't really scale your models.”

However, AIs can be trained on video, audio and synthetic data, and companies seem poised to extract personal data. Professors Angela Hui Zhang and S.Alex Yang warned in The Sunday Times that the GPT 4o “free” model is a way for OPN to store a lot of crowded multimodal data.

ChatGPT is good at intercepting zero-day vulnerabilities.

A few months ago, a group of researchers revealed that revealing security vulnerabilities to the GPT-4 agent enabled it to hack a series of test websites. But while it's good at attacking known vulnerabilities, it performs poorly against unknown or ‘zero-day' vulnerabilities.

The same researchers employed a team of sub-agents to lead the GPT-4 scheme agent to try and identify unknown or zero-day vulnerabilities on test sites. In a new study, AI agents were able to exploit 53% of zero-day security vulnerability opportunities on test websites.

Luma AI's new Sora competitor, Dream Machine

Luma AI dropped the new Dream Machine text and image into a video generator, and the usual AI influencers put out some amazing high-quality long threads.

Unlike Sora, the people can try themselves. On Reddit, users have reported that it's taking an hour or two to produce anything (due to overloaded servers) and their results don't match the time.

500 thousand new AI millionaires

If you're not a millionaire from the AI ​​boom yet, are you trying? According to consulting firm Capgemini, the total number of millionaires in the US rose by 500,000 people to 7.4 million. Fortune likens it to the AI ​​stock boom.

The publication notes that investors' optimism in AI has driven the S&P 500 up 24% over the past year, Tesla has doubled, Meta has jumped 194%, and Nvidia has surged 239%. The index and the tech-heavy Nasdaq have hit record highs this year. Growth looks set to continue, with Goldman Sachs predicting that AI investment globally will top $200 billion by 2025.

Also read

Main characteristics

Bitcoin: A Peer-to-Peer Online Gambling Payment System by Satoshi Nakamoto

Main characteristics

Lost or Lost: Can Blockchain Protect the Heritage of Endangered Populations?

Adobe and Microsoft return to AI features

The company changed course following backlash from users over Adobe's terms of service, which gave the company broad permission to access and use user content and train AI. Now, Generative AI is saying that it will “never” train on creators' content, “or we wouldn't have thought of any of these exercises.”

The controversy began after content creators were told they had to agree to Adobe's terms or face a fee equal to 50% of the remaining annual subscription cost.

Microsoft has retired the recall feature on i-branded Copilot+ PCs. The device creates a screenshot record of everything users do to help AI understand it. But cybersecurity experts say databases are a big honeypot for hackers. The device now turns off by default, requires biometrics to access, and data is encrypted when users aren't logged in.

Andrew FentonAndrew Fenton

Andrew Fenton

Based in Melbourne, Andrew Fenton is a journalist and editor covering cryptocurrency and blockchain. He has worked as a film journalist for News Corp Australia, SA Wind and national entertainment writer for Melbourne Weekly.



Leave a Reply

Pin It on Pinterest