1. GPTBot, GPT-5 and other OpenAI changes

OpenAI showcased its web crawler - GPTBot. This raises some interesting questions.

First, what is the incentive for sites to allow GPTBot?

If you allow Google to crawl your site, they will cite the website in search results. That helps you get organic visitors to your site.

OpenAI on the other hand will never cite sources.Sure, there is a collective value from the crawled information. But for a website there is no incentive as they cannot get traffic. So, what will be the incentive going forward?

Secondly, now that the crawler is identified there is a high chance of bad actors messing with data going into the model.

We know people are using ChatGPT as a search engine. For example, it tells me the Ducky One 2 Mini is the best mechanical keyboard. What is stopping the manufacturer of lower ranked keyboard manufacturers from feeding bad data to move their rankings?

OpenAI is also rolling out some cool changes in the ChatGPT interface:

1. If you are Plus users you get:

   1. GPT-4 model by default.

   2. You can now upload multiple files to Code interpreter and generate insight across multiple files.

2. Prompt examples - Previous examples were terrible. OpenAI is now promising better opening examples.

3. Suggested replies to deep dive into the topics

4. Keyboard shortcuts.

The feature I am most looking forward to testing is multiple upload to Code Interpreter.

Quite a lot of data is normalized. That is there is only a single instance. For example, a customer database doesn’t repeat customer names across multiple tables. It has multiple tables like - customer master table, customer address table, customer sales table etc.

When you want to do a customer analysis you first denormalize the data. That helps increase read performance and avoids costly joins in a normalized, relationship database.

I am curious to see how Code interpreter works in the multi file scenario. Can it create denormalized data and run analysis? .

Additionally, OpenAI filed for a copyright trademark for GPT-5. One can always speculate if OpenAI is now preparing for GPT-5 training/release.

2. AudioCraft - Generative AI for Audio

Meta released Audiocraft - generative tool for Audio. The tool contains three different models:

1. MusicGen - Meta’s text to music model. This has been trained on Meta’s own music library.

2. AudioGen - Meta’s text to audio model. It is different from MusicGen in that this can be used to generate normal audio. For example, dog barking, steps on a wooden floor etc.

3. EnCodec - This is the decoder to allow for better music generation.

The generated samples in the page are quite good. So, check it out.

Also, touching upon the whole “open” source part of the announcement. In this case you can look and audit the code for research purposes. There are no commercial licenses.

To me it sounds like Meta is laying ground for whenever it releases an AudioCraft based tool. AI results are often black boxes. No one knows why it generates the final results. But the “open” code will allow for a clean audit of the results.

This is in contrast to OpenAI’s approach which has not released neither the research nor the code for GPT-4. That makes GPT-4 a black box for people. No one knows how those results are being generated.

“Open” source and auditable results might be the future of generative AI tools.

3. LLMs and AI assistants

This week we saw interesting developments in the AI assistants space.

First, Microsoft is shutting down Cortana. The company is focusing on modern AI advances, including Bing Chat and other AI-powered features. Those features will be powered by Microsoft's partnership with OpenAI. The new ChatGPT-based Bing experience will be integrated into Windows 11.

Cortana will no longer be available as a standalone app in Windows. Microsoft provides alternative options for voice access and productivity needs, including Windows 11 voice access, the new AI-powered Bing, Microsoft 365 Copilot, and Windows Copilot.

Google is planning to overhaul the Google Assistant. The focus now will be to explore the potential of generative AI using the latest LLM technology. Google will start with the mobile version of the product.

The two opposing approaches are interesting. Cortana is a well known brand so Microsoft rebranding and going the Google Assistant way made more sense. But they have decided to re-do things.

🗞️10 AI news highlights and interesting reads

1. This is my favorite read from last week. If you are building an LLM product you might want to read about the patterns in LLM based products.

2. One of the annoying ChatGPT behaviors is that it apologizes. A lot. How do you get ChatGPT to stop apologizing?

3. Impending death of another “summary” based tool. The Youtube summarizer. Youtube is working on its own tool. 

4. Zoom has changed the ToS to allow it to use customer data to train AI and there is no opt-out. (This is a PSA and the link is to the ToS. Read 10.2).

5. Github co-pilot will now show you if the code matches something from a public repo. That should help people decide if they are infringing on any license by using such code.

6. AI might not replace humans - but humans with AI will replace humans without AI. To be honest I am skeptical of this piece but I think it provides an alternate way to look at the generative AI tools.

7. What is up with people creating “persona” chatbots? It seems like a waste of generative AI technology. Meta is now on this boat too.

8. AI investment to approach 25 billion by 2025. But Investors are not getting swept by the AI hype. 

9. News Corp in Australia is using AI to generate 3000 articles per week. Medium goes the other way and says human writing will be paid more.

10. Silicon Valley and Pentagon - a match made in heaven? 

1. Weird world of LLMs.

2. The Impact of ChatGPT and other LLMs - A series of videos from MIT Physics department featuring speakers like Prof. Yann LeCun and Dr Stephen Wolfram.

3. Practical AI for Teachers and Students