Synapses#16 — Last Week in GenAI

Last week sent shock-waves through GenAI with the sacking of Sam Altman. How he was fired so easily? And more on the best 7B model and visual instruction tuning.

| Weekly Edition- November 20, 2023 |

Latest top 3 impactful advancements in Generative AI, across research, product innovations, and business.

Chart of the Week

Does spending more money on healthcare make it better?

In general, there is a trend line which says that countries who spend more on healthcare does imporve quality of life; but with one big exception, USA.

USA spend more than anyone but still has quality of health as third word country. Why is it?

• Maybe because govt. incentivizes over-charging by insurance & pharma companies
• Maybe because the cost of drug innovation has to be paid
• Maybe because migrant influx can draw the average down

Maybe because our healthcare system simply sucks!

1. Research

Visual Instruction Tuning

Common approaches for tuning visual instructions in current methods usually involve presenting extensive textual descriptions to large language models, prompting them to generate instruction-following data. Although these methods have demonstrated impressive performance, the textual descriptions they rely on are often derived from coarse-grained image annotations. Additionally, these instructions may inadvertently conflict with the visual content when the entire visual context is not taken into account.

Credit Paper

To overcome this challenge, a new set LVIS-Instruct4V, a dataset of fine-grained visual instructions. This dataset comprises 220,000 visually aligned and context-aware instructions generated by employing the powerful GPT-4V model, prompted with images from LVIS.

Paper-

To See is to Believe: Prompting GPT-4V for Better Visual Instruction Tuning

2. Product

Mistral-7B

When a three-founder team secured stratospheric seed funding, I was skeptical. However, the Mistral team has delivered a truly impressive model in record time, something that big tech is finding challenging to achieve. It underscores the point that AI is in the hands of a select few who truly understand the craft.

Their 7B model doesn’t compare to most in the category.

Credit- Mistral

Why it matters:

• The model is under Apache 2.0 License.
• The 7B is the most important variant for LLMs, as it can be accommodated on a single GPU (crucial for GPU crunch).
• Moreover, the 7B is the most sought-after model category for enterprise LLMs to keep inference costs down.”

3. Business

OpenAI Coup

On Friday my whatsapp buzzed about Sam Altman’s firing as OpenAI CEO. While I dont know why, I can probably explain How?

Credit- CNBC

Why was it so easy for the board to dismiss OpenAI CEO Sam Altman and Chairman Greg Brockman? And why did they do it when the company is flourishing, boasting a waiting list for GPT-4 and a 1000x growth in valuation?

In a typical private firm, firing a co-founder is not an easy task, given their significant stake in the company. Plus, solid reasons are usually required to dismiss leadership in a hyper-growth company. As it turns out, OpenAI is not a typical company. OpenAI started as a non-profit, with its stated goal not to make money (funny ha). This is reflected in its name, “OpenAI,” indicating its initial aim was to conduct “open” research to build AGI. (even Elon Musk was part of it then). Despite abandoning this goal, the board is still accountable for it, at least on paper. This can be used as an excuse for the ouster, even though the company has been running the same for-profit product for the past year.

Furthermore, it implies that Sam doesn’t hold any equity in his company, giving him less leverage against disgruntled board members. A non-profit board often resembles a housing society board. OpenAI board has fiduciary responsibility but not to its shareholders ; rather to its own charter like a trust.

I could not find any news sharing details of the term sheet when major investors like Microsoft (47%), Thrive, and Khosla invested in it. It would matter if they have observer status on the board or if information rights or some outstanding warrants.

There’s a possibility of a counter-coup against the OpenAI board from its major investors. However, to execute this, they would need to separate the for-profit section of the company from the non-profit, which may not be an easy task.

When I was in grad school, I aspired to start a tech non-profit. However, all the lawyers I consulted strongly advised against it, citing issues about intellectual property. They asserted that in every case, it ends up being a painful separation. According to what I was told then, a non-profit has to agree to sell all its assets to a for-profit. Since OpenAI has backing from big PE firms, money may not be the primary issue, but ego could be. Nevertheless, it’s likely that there could be a behind-the-scenes deal to resolve this issue. After all, it’s a $100B company; no one is going to let it go easily.

In a nutshell, you haven’t seen the last of this boardroom drama. Stay tuned for the next episode in this Game of Thrones!

Weekly Wisdom

The first principle is that you must not fool yourself and you are the easiest person to fool.

That’s all you need to know this week. Keep Generating new ideas with Generative AI …..

Follow Towards Generative AI for the latest advancements in AI.