Time to Snap Out of Hallucinations

Generative AI chatbots behave like know-it-alls. Omniscient, maybe, but they can also be sporadically unreliable with answers. At times, they enter a dream-like, zone-out state and start hallucinating.

That’s what went down at the event when Google released its much-anticipated chatbot, Bard. In a promotional video, Bard was asked about new discoveries from the James Webb Space Telescope. The bot gave three answers and also said JWST took the first picture of a planet outside of our solar system. Wrong answer!

However, scientists look at these errors, also called hallucinations, a little differently. They think of hallucinations as an innate part of large language models, a byproduct of the way they compress their inputs, losing track of factual relations in the process. The biggest issue is not with the problem itself but with its perpetuity. So, have we hit a dead-end here?

Not really. Retrieval augmented language modelling, also called?REALM, has been suggested by experts as a redressal to the hallucination problem with LLMs. The REALM language models utilise external sources to train on data. It has a ‘knowledge retriever’ to search the document that will probably have the information relevant to the prompt.

When a user inputs a prompt such as "Vincent Van Gogh was born in”, a conventional LLM will attempt to predict the next token in the sequence to complete the sentence. The model can then output Van Gogh’s birthplace from maybe a Wikipedia page and use this to generate a more reliable response. The knowledge retriever is also able to produce references to the knowledge documents, which helps the user verify the source and accuracy of the text that the model generated.?

While there’s a huge discussion about the hallucination of chatbots, there’s another aspect to it. Google, ChatGPT and others are making efforts to mitigate the hallucinations, but, isn't it true that hallucinations are a crucial aspect of intelligence??

Read the full story?here.

Google Graveyard Grows

In 2017, Google was on an acquisition spree. It acquired four platforms from Twitter such as Fastlane, Crashlytics, Fabric, and Digits. Fastlane helped Google and companies alike in automating building, testing, and deploying applications – run with a single command. But now, Google seems to be least concerned about it.

Six years later, Fastlane is almost dead. Google has abandoned the application and has not sponsored it since the end of November 2021. And though Google still owns it, the community “does all the work”. If Google decides to kill the application, it will be one more addition in the wide field of Google’s graveyard.?

Read the full story?here.

End of Web Search

The new generative AI-powered assistance tools are going to transform the three-decade-old search engine method. After users witnessed the power of ChatGPT, which gives precise answers, there seems to be no turning back for them to explore search engines that provide a never-ending list of articles from which one manually hunts for the answer. The approach is inefficient as most of us don’t even visit Google’s Page 2.?

Read the full story?here.

VCs Charge EV Space

The electric vehicle (EV) market has witnessed a significant influx of investments worth $76.2 million in just the first two months of 2023, and it shows no signs of slowing down. In 2022, the Indian EV sector attracted a total funding of $1.66 billion, marking a 117% increase from the previous year's $766 million, effectively valuing the market at $3.21 billion.

The investment environment is also charged with government regulations, which are providing subsidies for EVs. Seeing the government enthusiastic about it, LPs have also focussed towards climate-related projects, especially EVs.

Read the full story?here.