Suggestions
Duncan Anderson
Founder @ Humanise.AI
Duncan Anderson is a prominent figure in the field of artificial intelligence and technology, known for his role as the Co-founder and CEO of Humanise.AI. He has a rich background in AI, having previously served as the European CTO for IBM Watson, where he was instrumental in shaping the technical strategy of IBM’s AI initiatives across Europe. His experience includes engaging with both executive-level clients and developers, enhancing AI implementations for various industries, including finance and healthcare.12
At Humanise.AI, founded in April 2017, Anderson focuses on innovative applications of Conversational AI to enhance customer engagement and automate business processes. The company's platform aims to facilitate seamless interactions between customers and businesses, allowing for human-like conversations through AI while enabling human intervention when necessary.14 Under his leadership, Humanise.AI has developed solutions that streamline customer journeys across multiple communication channels, such as chat, SMS, and email.2
In addition to his work at Humanise.AI, Anderson is also the Founder of Barnacle Labs, which he established to help organizations leverage generative AI technologies effectively. His expertise extends beyond technical roles; he is actively involved in consulting and sales strategies within his ventures.2
Anderson's educational background includes a 1st Class BSc (Hons) from the University of Hertfordshire. He has received several accolades throughout his career, including the "Best of IBM 2015" award, recognizing him as one of the top performers among IBM's extensive workforce.2 His thought leadership in AI is further evidenced by his participation in industry conferences and contributions to discussions on emerging technologies.3
Highlights
I’ve had a couple of people ask me recently how much it costs to train an AI model.
👉 Here’s my answer: it depends.
Here’s why: 1️⃣ Maybe you are fine-tuning an existing small model on your MacBook using Apple’s MLX framework. In which case the answer is £0.
2️⃣ Maybe you are fine-tuning a larger model on cloud servers? Larger models still vary considerably in size, as does the nature and size of a training set. I’m going to say you might be in the “a few hundred to a few thousand, maybe a few tens of thousands”. And that’s a very ballpark figure, there will be exceptions outside of those bounds.
3️⃣ Maybe you are training a new LLM from scratch? In which case you’re definitely into the millions, even 10’s or possibly 100’s of millions at the extreme. It varies a lot depending model and dataset size. Elon Musk just raised $6bn to do this, so the sky really is the limit.
So there’s your answer. You can’t answer this question unless you know more about the scenario. My three scenarios aren’t exhaustive, by the way, merely illustrative.
To back this up, let me give you two examples:
- Bloomberg created BloombergGPT on 512 Nvidia A100 GPUs, each of which would set you back in the order of $10k - so $5m of servers.
- Databricks original Dolly LLM was created with a claimed $30 of compute finetuning time.
So it depends.
If someone insists on a more specific answer without defining a specific scenario, you should probably walk away from the conversation. Demanding specific answers to vague questions is a warning sign. I might as well demand an answer to “how much does it cost to build a software application?” I can ask, but I’m unlikely to get a useful answer. Thankfully, most people understand nuance and ambiguity!
This is a fascinating interview with one of the OpenAI founders, discussing their approach to openness of models.
Conventional thinking says that "open is best" - share the algorithms, research, etc. Let the community understand their biases, find their limits, be reassured about how they work, etc. Let researchers build on what's gone before. It's how most of machine learning has developed so far.
But... “We were wrong. Flat out, we were wrong. If you believe, as we do, that at some point, AI — AGI — is going to be extremely, unbelievably potent, then it just does not make sense to open-source. It is a bad idea... I fully expect that in a few years it’s going to be completely obvious to everyone that open-sourcing AI is just not wise.”
Why is it a bad idea?
- Why would OpenAI give their best work to all their competitors?
- Do we, as a society, want these tools in the hands of rogue states, terrorist organisations, etc?
Openness has brought many benefits, but perhaps with AI it might also bring dangers. No clear answers and some will no doubt disagree – but a fascinating debate.
https://lnkd.in/emzUnC_s
#machinelearning #ai #opensource #openness
