The Future of AI Integration

Redefining Multilingual Speech Technology, Microsoft's Integrated AI Plugins, and Lessons in AI Dependency

For this recap, we have:

  • 🔥Top 3 AI news in the past week

  • 🗞️10 AI news highlights and interesting reads

  • 🧑‍🎓3 Learning Resources

🔥Top 3 AI news in the past week

1. Expanding Language Horizons

Facebook has released an open source model called MMS (Massively Multilingual Search) for STT (speech to text), TTS (text to speech) and language identification.

This is a big breakthrough. Currently, STT and TTS models recognize only 100 languages. With this the technology has been expanded to 1100 languages. That is 10x the current best.

Additionally, these models can recognize 4000+ languages.

As per Facebook, they also have half the error rate of OpenAI’s Whisper.

These guys are on a roll.

2. Bing Chat Enters the OS

After Google’s announcement, it was time for Microsoft to announce AI products. Here’s a rundown of what was announced during Microsoft Build:

Windows Copilot - Microsoft is integrating Bing Chat into the OS. Now you can do everything you could do with Bing Chat, now on the OS and across all installed applications. You can do the usual stuff - summarize emails, documents, re-write etc. But it goes beyond that by integrating into the installed applications.

Microsoft is also adopting OpenAI's plugin model. So, you can use ChatGPT and Bing plugins to interact with the integrated AI.

The great thing about it is the direct integration into the OS. Eat your heart out, Mac users – at least for now 😀. Until Apple announces something similar. And someone will come up with an alternative solution. Especially, because of the privacy concerns with Microsoft telemetry.

The bad thing is - the security aspect of the plugins. It can open a whole new attack vector on the OS and antivirus softwares might struggle with it.

It also might be the second nail in the coffin for all the summarize, “talk to your document” apps. Once, this feature is integrated with Google Docs and Microsoft Office - why will you want to pay for extra apps?

Search comes to ChatGPT - Looks like OpenAI had enough of the testing and new features are being rolled out left and right.

No prizes for guessing the search engine behind it. Ding, Ding, Ding..It’s Bing!

Co-Pilot in PowerPages - Microsoft is now adding AI to their PowerPages platform, their low-code tool to build websites. It’ll help users to generate text, forms etc.

Microsoft Fabric - A new data analytics platform built on top of Azure Data lake but can get data from S3, Google cloud etc. It can help users build pipelines, write code and, build ML models.

3. From Trusted Advisor to Nightmare: The Hazards of Depending on AI

A man filed a personal injury lawsuit against Avianca airlines. Avianca's lawyers wasted no time and requested the judge to dismiss the case. The man's lawyer had a different plan in mind. He submitted a document citing half a dozen cases that bolstered his client's claims.

Here's the twist—the judge and Avianca's lawyer couldn't locate any of the referenced cases. Quite a conundrum, right? The lawyer was then asked to provide copies of these elusive cases. The lawyer submitted screenshots as evidence, taking extra precautions to ensure their authenticity.

You already know the direction this story is taking.

The lawyer had used ChatGPT to compose his brief. But little did he know that ChatGPT had supplied him with fake cases.

When asked to file tangible copies of these cases, the lawyer turned to ChatGPT once again. ChatGPT had reassured him that the cases were genuine. Feeling emboldened, the lawyer used ChatGPT to provide the requested copies. He even went as far as incorporating chat screenshots into a legal document.

The lawyer maintains that it was never his intention to deceive the court. He expressed regret for relying on ChatGPT for their research. Unfortunately, the judge isn't pleased with this turn of events. The judge has threatened sanctions against both the lawyer and his firm.

It serves as a stark reminder of how ChatGPT has fooled many people. There is a clear warning stating that ChatGPT may produce inaccurate information. But many tend to overlook these warnings. Even legal professionals!!

This story carries significant importance for those who fear job insecurity. The lawyer and his firm could have prevented the entire debacle. They should've used paralegal services. They instead relied on ChatGPT's. It's a hard lesson learned the hard way.

My sincere hope is that this story serves as a valuable lesson. It helps people avoid making similar mistakes. The legal community might become apprehensive about ChatGPT's use moving forward.

🗞️10 AI news highlights and interesting reads

  1. OpenAI says in 10 years AI could be as productive as one of today’s large corporations. This poses an existential risk and they suggest some regulations to manage it. To achieve this, countries need to form something like the IAEA. The IAEA is an intergovernmental agency under the UN to oversee nuclear energy. This “AI agency” will monitor the AI systems and conduct inspections. Just like nuclear energy is tracked through signatures, they suggest using compute and energy usage to track systems.

  1. In the meantime, Google is working on voluntary rules until there are some real regulations in place.

  1. Sharing prompts and results has been a pain. Taking screenshots is one way. But then everyone has to type in the prompts manually. Or you can share as plain text. But ChatGPT results are non-deterministic. So, the results might not be the same. Even the lawyer above would’ve loved this feature. Now you will be able to share your ChatGPT conversations publicly.

  1. LLM Agents and plugins need to connect to tools to perform the tasks outside the LLM environment. So, it is important for the LLM to know which API to call and pass correct arguments. Gorilla is a fine-tuned Llama-model which can generate the correct call and arguments.

  1. If you are trying to build something beyond a document summarizer or a wrapper around GPT4 API, things can be hard. Finding the correct context window, dealing with slow responses (I am looking at you GPT-4) etc are some of the problems.

  1. Chatbot leaderboard for the week. GPT-4 is still ahead.

🧑‍🎓3 Learning Resources

  1. Build a product using Replit+AI. The author is a non-technical and won a hackathon competing with engineers.

  2. LangChain 101.

  3. NLP Course from HuggingFace

That’s it folks. Thank you for reading and have a great week ahead.