Discover more from AI SIMPLIFIED By Nerdynav
Are Smaller AI SaaS Obsolete? (Microsoft's AI Monopoly)
Stanford's Alpaca, AI in Gdocs & Sheets, Copilot 365, BritGPT
AI just got cheaper, accessible, and more ubiquitous. Then why do I say smaller AI companies like Jasper, copy.ai, etc. have become obsolete? Read to find out.
1. Stanford's Alpaca
Stanford researchers have unveiled an artificial intelligence (AI) model that performs almost at par with ChatGPT but cost them only $600 to train.
Alpaca is a seven-billion parameter variant of Meta’s LLaMA. It was fine-tuned using 52,000 instructions generated by GPT-3.5 (ChatGPT). (Much like the way, human testers were used to fine-tune ChatGPT, Stanford used the model behind ChatGPT to train their Alpaca AI.)
In tests, Alpaca performed comparably to OpenAI’s model but produced more hallucinations.
Alpaca is significant because it proved that building and training new AI models can be insanely cheap.
This could potentially enable more people, including bad actors, to create new cheap models.
It also shows once you make your model public, even without revealing its code, it can be used to build better models by competitors (like using ChatGPT to instruct Alpaca while training).
Could this make companies like OpenAI, Google, and Microsoft even more aggressive in guarding their proprietary tech?
On the plus side, the future where you will be able to train your own ChatGPT-like AI using your own computer just got closer.
2. AI In Google Workspace and Microsoft 365
Both companies have indicated that AI will be heavily incorporated into their apps. Google showed visuals of their AI being used in Gdocs to write full articles, in Sheets to write formulae, and in Slides to generate complete presentations with text and AI-generated images.
Microsoft's demo was better. They showed their AI tools live in action. Their Copilot AI will be available across apps like Word, Presentation, Excel, etc, and will be able to aggregate data about any topic across these apps to answer questions, plan meetings, generate responses, etc.
Copilot will also be able to take live notes in meetings and recap the discussion done so far.
While this means a massive boost to our productivity, it also spells the death of many SaaS companies built around providing GPT-based features in these apps. Even tools like Jasper, copy.ai, Merlin, etc. might be rendered obsolete.
All these apps provide a wrapper over one aspect of GPT like writing, e-mail replies automation, idea generation, etc. Microsoft Copilot and Google suite will do all this and more - for free. Appls providing AI/GPT in Google sheets will go the same way.
It also raises concerns about privacy. If Microsoft and Google collect even more data, in name of AI, do we really know how it will be used?
3. Bard's Limited Release
Google has started rolling out its AI chatbot Bard, but it is only available to certain users in US and UK and they have to be over the age of 18.
The initial response to Bard has been lukewarm with users complaining that is much inferior to Bing Chat.
It demonstrates less creativity and is prone to more mathematical errors. In fact, a button to "Google It" appears after every response of the bot, as if an indicator of its own insecurity.
It is seemingly faster than Bing but that's likely due to a smaller model. A smaller model would also explain its poor performance.
4. OpenAI Unveils GPT-4
GPT-4 is multimodal, meaning that it accepts both text and image inputs. It is better and safer than ChatGPT.
Some of its abilities shown in the demo:
GPT-4 can understand hand-drawn mockups and convert them into working website code.
It can analyze complex documents like tax codes but also perform accurate maths besides citing appropriate laws and principles to calculate tax.
GPT-4 also outperforms GPT 3.5 in human exams like Bar, SAT, and GRE (scoring in the 90th percentile vs ChatGPT's 10th percentile) and has a much longer contextual memory.
It’s only available in the ChatGPT Plus paid subscription and via an API waitlist.
Bing chat has also been using an early version of GPT-4 for the past 5 weeks. BeMyEyes (a startup for the visually impaired) and Khan Academy became one of the first users of GPT-4.
I am curious about the results we could get if Alpaca is trained using a bigger LLama model and fine-tuned with GPT-4.
5. Global AI Race Begins
The UK government is investing £900m in the supercomputer as part of an artificial intelligence strategy that includes ensuring the country can build its own “BritGPT”. The aim is to counter China’s AI influence and to ensure that the UK remains competitive in the field of AI.
6. Google's PaLM and Makersuite
Google Cloud has announced that its artificial intelligence (AI)-powered applications, such as the Pathways Language Model (PaLM) API for language models and the Makersuite prototyping tool, are now available to developers.
With MakerSuite, developers can iterate on prompts, augment their dataset with synthetic data, and easily tune custom models.
7. Baidu's Ernie
Ernie is an AI model developed by Baidu, a Chinese multinational technology company.
Ernie is likely to enjoy a significant market advantage on its home turf over U.S.-made products, due to both China’s own drive for technological self-sufficiency and the country’s strict internet censorship.
However, its demo was a lackluster presentation with pre-recorded responses.
8. Apple's Foray Into LLMs
Apple recently conducted an internal event focused on generative AI content and large language models (LLM), which are the neural networks that power chatbots like ChatGPT.
Apple has a huge cash reserve and with its track record of delivering highly intuitive products, it will be exciting to see what they do in the field of AI.
9. Midjourney Releases V5
An AI-generated "synthetic photograph" of a girl through a window generated using Midjourney v5 by Julie Wieland.
Midjourney v5 brings with it improved “efficiency, coherency, and quality.”
V5 now responds with a “much wider stylistic range” than version 4, while also being more sensitive to prompting, generating less unwanted text, and offering a 2x increase in image resolution.
Midjourney v5 can generate realistic human hands fairly well, which was a problem with previous versions.
It also generated believable portraits of humans and people in natural poses. Will it take opportunities away from small-time models and graphic designers?
10. Microsoft's Bing AI Image Creator
Microsoft has given its AI image generator its own dedicated Bing Create site. Bing Image Creator is powered by an advanced version of OpenAI’s DALL-E model and performs surprisingly well even with natural language inputs.
The images are free to create and the more descriptive you are, the better output you get.
AI Business Insights
AI Business Opportunity for Storytellers
There's potential for new startups to come up in this space that uses AI-generated stories, and combine it with AI visuals, and AI voice-overs for realistic and personalized storytelling experiences.
Featured AI Startup: Oscarstories 🦊
Curated AI Tools For This Week 🔥
Every week, I will be recommending 3 AI tools that I have tested and believe to be useful for my readers.
I want to promote free, open-source AI tools so I will try to ensure that the list has at least 1 open-source AI tool available for free.
If a paid AI tool I endorse has an affiliate program, I will also include its affiliate link which earns me a referral commission. It does not affect what I pick and is just a way to support my blog without ads.
I’d love to hear from you too! You can nominate your favorite AI tools for next week’s list by mentioning them in the comments.
Here are my picks for this week:
Sudowrite is an AI story writing assistant with many helpful features like Rewrite, Describe, Brainstorm, Twists, and Character generator.
Created by a duo who are sci-writers themselves, it is a lot of fun to use and the stories generated by Sudowrite feel surprisingly human. Read the full review with example stories. Or, try it for free.
Murf is an AI voiceover creator. A great collection of AI voices that sound human (with the best African-American voices out of all AIs tested).
Perfect for Podcasts, e-learning, and ads. Vary emotion, pitch, speed, emphasis, edit pronunciations, and more. You can also clone your own voice or swap existing voices with clearer voiceovers. Read review. Or, listen to Murf's voice samples.
3. Modelscope Text To Video
You can now create amazing videos from text with the largest diffusion model ever open-sourced. Modelscope Text-to-video is a completely free AI text-to-video tool that you can use right now to create AI videos from text prompts. Access it by visiting this link.
I know this newsletter is very long but I hope you found it useful and interesting. Let me know your feedback and thoughts in the comments.
Have an amazing day. And, stay awesome! 😊
Sources for Images: ARK, Microsoft, Google, OpenAI, Murf, Sudowrite, Oscarstories