This is directly from The WSJ linked above.
It is just various paragraphs taken. It will only make sense to read the article in its entirety.
The inner workings of frontier artificial-intelligence models from Google, OpenAI, Anthropic and their competitors are trade secrets worth billions of dollars. So it was big news when the source code for Anthropic’s crown jewel Claude Code unintentionally surfaced. It cracked open a system that made Anthropic’s offerings a Silicon Valley favorite and put the company on a path to beat ChatGPT-parent OpenAI to profitability.
The leaked code doesn’t let people make a copy of Claude. But it revealed enough to be a useful illustration of what’s changed since the debut of ChatGPT. Here are three key elements that make today’s AIs far more useful and reliable than they were even a year ago:
AIs now know more—and can look up what they don’t.
When the LLMs, or large language models, that power generative AI were in their early days, developers trained them on available digital media—books, websites, social posts, videos. Now real people are generating data just for them. In what was once a cottage industry, startups worth billions now pay humans to distill their hard-won subject-area expertise into lessons for the bots.
And the bots we interact with have a separate trick borrowed from humans: They look things up on Google or some other search engine. This has led to another burgeoning industry in scraping Google’s results and delivering them to other AI companies, so they can rapidly deliver the internet’s most current knowledge.
OpenAI is public about the work it has done in both enhancing models’ knowledge and making them better at looking stuff up. In the two years since the company released its GPT-4o model, internal tests show that its latest main model issues 26% fewer factual errors.
Anthropic’s co-founder has said the number one thing customers ask for is that chatbots be more honest and free of hallucinations. To that end, the company is researching why models confidently free-associate, and uses techniques during training to try to get them to more often admit the gaps in their knowledge.
“Where Claude consistently stands out in independent evaluations is what researchers call ‘calibration’: knowing what it doesn’t know, and saying so,” says an Anthropic spokesman.
AIs are growing adept at using tools.
[This is huge -- how the writer said / wrote this -- anthropologists -- when did a primate become human -- when it began to use tools.]
A gnarly debate at the center of modern AI research boils down to this: Can a system modeled on the human nervous system ever actually match a human’s abilities? So far, the answer has been “no.” That’s because people don’t just guess at answers. Since the dawn of writing, we’ve used a little trick called symbolic reasoning, aka math. Fun fact: Humanity’s first known scrawls were made by accountants.
Generative-AI models early on would suggest likely answers to math questions. Now, they can use real math. They recognize requests for calculation and either access an available software tool, or write their own code to solve the problem. They’re falling back on traditional, run-of-the-mill calculators.
AIs now check their own work—and each other’s.
This really caught my attention!
In the beginning, chatbots spewed answers in a stream of not-quite-consciousness. Now, unless we ask a very simple question, the AI chatbot performs a “chain of thought”: The AI has a conversation with itself to arrive at a suitable answer. Some bots go further, by asking a different AI model—usually, variants of themselves—to gut-check an answer.
This is simply amazing.
Now, when companies develop AI systems for their own specialized needs, they can opt to have the results generated from one AI run by a model from a different provider altogether—say Claude checking ChatGPT. The answer is only deemed acceptable if both AIs agree on it, says Pavel Kirillov, chief technology officer of NineTwoThree, a consulting firm that builds AI-based systems for clients ranging from FanDuel to Consumer Reports.Kirillov calls this approach a “council of models,” and he says the results are better quality, with lower error rates.
Today’s underlying AI models are smarter than they were a few years back, but the AI services they power are more effective because they use fresher information, traditional software—and each other.
Transcribing the above and listenint to YouTube music suggests that music is going to be just one more huge universe for AI.
Google owns YouTube. If one pays attention, one can see that Google is using AI to curate a song list. I don't have the time to explain, but all you have to do is ask AI the "right musical prompt," read the chatbot's reply, and then scientifically observe what YouTube hands off to you. This is not random. YouTube is incredibly sophisticated, and Steve Jobs anticipated it as far back as the iPod and then the first iPhone. If you need proof, listen to his keynote speech when he introduced the world to the iPhone.
No comments:
Post a Comment