Rafie is back! In this episode, we delve into Multilingual AI and AI-enabled translation services, machine learning translation techniques, and we discuss and compare translation tools such as Google AutoML, GPT-4, Marian NMT, DeepL, Welocalize, Llama 2, Mistral and BLOOM
Furthermore, we discuss the potential and impact of multilingual AI for global startups (the andreesen horowitz view is that the company of the future is 'default global').
06:00 - Machine Translation vs LLMs
16:10 - Where GPT-4 and LLMs perform badly
29:10 - Mistral Model Weights Leak
We discuss: Multilingual AI, Machine Translation, Parallel Data, Marian NMT, DeepL, Welocalize, WMT, GPT-4, Translating Prompts, Multilingual Utterances, Localization, Prompt Pipelines, GPT-5, Bloom, Parameters, Mistral, LLMs, MT Engines, Token use
Links:
- https://marian-nmt.github.io/
- https://www.deepl.com/en/translator
- https://www.welocalize.com/do-llms-or-mt-engines-perform-translation-better/
- https://machinetranslate.org/wmt
- https://machinetranslate.org/wmt23#czech--ukrainian
- https://mistral.ai/news/mixtral-of-experts/
- https://cloud.google.com/translate
- https://www.reddit.com/r/MachineLearning/comments/1452ziq/d_llms_in_languages_other_than_english/
- https://instruct-multilingual-frontend-dtjnk4f6ra-ue.a.run.app/
- https://heidloff.net/article/llm-languages-german/
- https://www.theverge.com/2022/11/2/23434360/google-1000-languages-initiative-ai-llm-research-project
- https://www.theverge.com/2022/7/6/23194241/meta-facebook-ai-universal-translation-project-no-language-left-behind-open-source-model
We mentioned as a possible example the idea of translating vs locally-generating a French employment contract in France. This was in relation to our AI legaltech startup https://www.genieai.co
**Watch Using AI on YouTube (and see our daft AI-generated background images)**: https://www.youtube.com/@genieai
**Share the show with friends and family using the all-platforms show link here:** https://podcasters.spotify.com/pod/show/using-ai
Please follow the show, and leave us a review when you get a chance!
Thanks,
Alex, Alex and Rafie
Two more notable mentions in our AI market leaders series: Perplexity makes a name for itself by aggregating famous models from major entities like Google, Mistral, Anthropic, and OpenAI, posing a new alternative to giants like ChatGPT and Google search. NVIDIA's Nemo is a model-building toolkit that allows on-premise AI deployment and is hidden behind an NDA while in early access.
We discuss: Perplexity, Rabbit R1, NVIDIA's Nemo, Artificial Intelligence, Large Language Models, Chatbots, Google Search, E-commerce, Machine Learning, AI Hardware, On-premise Deployment, Open AI, ChatGPT
Links:
- TechTarget News on Perplexity's Funding https://www.techtarget.com/searchenterpriseai/news/366565352/Perplexity-AI-secures-736-million-more-for-AI-search
- Perplexity's Official Website https://www.perplexity.ai/
- NVIDIA's Nemo Developer Page https://developer.nvidia.com/nemo-llm-service-early-access
- 75 Nemo Models on Hugging Face Model Hub https://huggingface.co/models?sort=trending&search=nemo
**Watch Using AI on YouTube (and see our daft AI-generated background images)**:
https://www.youtube.com/channel/UCHsQu4IipA7Ri2AqKcQZ1Yw
Two startups have boldly stepped into the ring with OpenAI - Cohere and Mistral. We delve into Cohere's enterprise-focused approach, backed by a rumoured $1bn funding round, and Mistral's pioneering efforts with open-source models.
We also look at the broader landscape, pondering the all-important question: can these startups truly contend with market leaders in the AI space?
We discuss: Cohere, Mistral, Transformer Architecture, Chatbots, Enterprise AI, Open-Source Models, OpenAI competition, The Bitter Lesson, Richard Sutton,
Links:
- Cohere's rumoured $1bn funding round: https://www.inc.com/sam-blum/what-coheres-possible-1-billion-investment-signals-for-ai-startups-in-2024.html
- Cohere's CEO is the co-author of foundational paper "Attention Is All You Need": https://en.wikipedia.org/wiki/Attention_Is_All_You_Need
- Cohere's official website: https://cohere.com/
- Mistral's official website: https://mistral.ai/
- Mistral's models on Hugging Face: https://huggingface.co/models?sort=trending&search=mistral
**Watch Using AI on YouTube (and see our daft AI-generated background images)**:
https://www.youtube.com/channel/UCHsQu4IipA7Ri2AqKcQZ1Yw
We discuss: AI Tool, HuggingChat, AI startup valuations, Open Source Machine Learning, Transformers Library, LLM Repository, Model Library, Git, Github, Model Hub, Llama 2, Launching AI models, business moat, deploying AI models, AI in the cloud, AI model leaderboards, Image classification, language models, sentiment analysis, ai music, nocode AI, low-code AI, model space, AI community, AI ecosystem, ML Ops, Emergent capabilities, generalist models, specialist models, GPT-3.5, GPT-4, ChatGPT, AI Exploration
Links: - Competitors showing increased enquiries following the leadership Farce at Open AI with Sam Altman https://www.cnbc.com/2023/11/28/openai-competitors-hugging-face-and-cohere-report-increased-inquiries.html
API Token issues https://www.theregister.com/2023/12/04/exposedhuggingfaceapitokens/
Interesting Hacker News post on hugging face (with ex-HF workers weighing in on strategy): https://news.ycombinator.com/item?id=37248895
Reddit post: Does Hugging Face do too many things?: https://www.reddit.com/r/MachineLearning/comments/160ts9g/d_is_it_me_or_huggingface_do_too_many_things/
transformers library: https://github.com/huggingface/transformers
In this episode of Using AI, we delve into the capabilities and offerings of Hugging Face, the leading AI repository often likened to GitHub for AI enthusiasts. Hosting over 250,000 datasets and 500,000 AI models, Hugging Face has revolutionised the AI world with its open-source initiatives.
Watch Using AI on YouTube (and see our daft AI-generated background images): https://www.youtube.com/channel/UCHsQu4IipA7Ri2AqKcQZ1Yw
On this episode: Llama, AI models, Weights Leakage, Comparison, Usage, Meta vs Open AI, Data management by Meta, Open Source AI, AI research groups, Instruction, fine-tuning, RLHF, Hosting Llama, AWS Bedrock, Azure AI Models, local finetuning, and guardrails, ChatGPT, GPT-4, GPT-3, GPT-3.5
Watch Using AI on YouTube (and see our daft AI-generated background images):
https://www.youtube.com/@genieai
Links:
Welcome to another episode of Using AI. I'm your usual host, Alex Denne, and today, I'm accompanied by Alex Pap and Nitish Mutha (Founder of Legaltech Genie AI).
We start by introducing Llama and discuss its weights leaking incident. We also elaborate how Llama compares to other AI models and explain how to use it. The conversation takes a turn towards the Meta vs Open AI dispute, shedding light on their differences and impact in this space. We also discuss Meta's data management and how it can actually come up trumps on both privacy strategy here, and non-copyrighted multi-lingual training data.
We don't delve too deep into the already covered demo-gate scandal, don't worry! This episode features insights from Senior ML Research Scientist Alex Pap and AI Startup Founder and CTO Nitish Mutha.
We discuss: Google vs OpenAI for the long-term. GPT4 Vision, GPT5, Multimodal AI, Gemini Ultra, Gemini Pro, Gemini Nano, OpenAI Whisper, DALL·E 3, Chain of thought, Google, OpenAI, Bard, AI technology, Machine Learning
Welcome to Episode 17: and the 3rd episode in our AI Market Leaders mini-series - focusing on Google vs OpenAI. This episode dives into all the details of the release in Gemini’s Ultra, Pro, and Nano (and how that affects Alphacode2, and Bard). We also delve into multi-modal technology and its promise for the future.
Watch This Episode of Using AI on youtube
https://www.youtube.com/channel/UCHsQu4IipA7Ri2AqKcQZ1Yw
Topics Discussed:
Additional Resources:
Alex and Alex discuss the use cases for OpenAI tools such as GPT-4, compared to Anthropic's Claude 2. We also offer some sneaky insights on Anthropic and their roadmap.
Topics discussed: AI Startup Competition, Anthropic reportedly working on 1 million context window, ensemble LLMs, cohere, london, san francisco, Dario Amodei, Hallucination, Refusing to answer, API Tooling, Claude 1, Claude 2, Sam Altman
Kebab challenge images on reddit: https://www.reddit.com/r/midjourney/comments/183hv0m/the_forgotten_kebab_challenge_1979/
Using AI on youtube - see what Alex Pap was laughing so hard at at the start of this episode https://www.youtube.com/@genieai/podcasts
Links:
It's been a year since ChatGPT was released (yeah, feels longer than that doesn't it?). We’re going to look through it’s achievements, and then run a few episodes taking a closer look at the competition hot on OpenAI's heels
So this will be the first in a short 5-part mini series where we’re talking about the alternatives to Open AI - given the Fiasco which has dominated headlines for the past 3 weeks - details of which are still emerging.
Some stats on ChatGPT from the past 12 months
Using AI on youtube - see what Alex Pap was laughing so hard at at the start of this episode https://www.youtube.com/@genieai/podcasts
Topics discussed: GPT store, Sam Altman, Dario Amodei, Anthropic, Elon Musk, Greg Brockman, Darth Vader, Fired by the board, Open AI CEO
Links:
We discuss: AI Coding Tools, Code-Interpreter, Python, Regex, Network errors, Machine Learning, ChatGPT4, HTML Parsing, Github Copilot vs GPT-4
In this episode, we delve into a fascinating experiment where I, Alex Denne, under the watchful guidance of ML Research Scientist Alex Pap, try to get AI to writing some regex that can be run locally on my machine using python, on millions of documents.
The goal? To extract matching text from millions of HTML files.
It all inadvertently unfolds into an intriguing journey of trial and error.
For the no-code listeners, this episode offers first-hand insights into the application and limitations of AI coding tools and code interpreters (and why, for now, you probably still need technical help like Alex D did!)
At the outset, we were greeted by a seemingly promising result - a neat CSV file with the right column names but no entries as the AI successfully claimed to extract definitions only to produce an empty result.
In an attempt to further probe, the AI was prompted to read the first 100 characters for potential matches. Alas! In lieu of any found matches, it concluded the document must be lengthy and gracefully tapped out.
In addition, we had to deal with several network errors that may be attributed to the reported DDoS attacks on OpenAI.
After multiple hits and misses, we decided to start afresh with a new approach. We didn't exactly strike gold, but we learned a lot.
Through this episode, we touch upon topics like ChatGPT4 and the wonderful feature of 'dragging and dropping' files into GPT-4 Turbo.
Watch USING AI on youtube:
We discuss: Microsoft, Microsoft Copilot, Copilot for Azure, AI technologies, Microsoft Ignite 2023, CRM Assistants, CRM integrations, Generative AI for Enterprises, GPT Store, Open AI, Word Add-Ons, Excel, Dynamics 365, Copilot Studio, Satya Nadella, Sam Altman, Notion AI, Notion's New Q&A feature, What Powers Notion AI (Spoiler: It's GPT-4).
No Rafie today - but we press on. In this episode, we discuss the new offerings from Microsoft concerning its Copilot series, which uses generative AI technologies. During the Microsoft Ignite 2023 conference, the company introduced the Copilot for Azure, Copilot for Service, and Copilot in Dynamics 365 Guides. We also discuss the launch of Copilot Studio, a platform that offers tools to connect Copilot to third-party data.
The Copilot for Azure, similar to Google's recently announced Duet AI in Google Cloud, takes the form of a chat-driven assistant for cloud customers, helping with configuration suggestions and troubleshooting potential issues. Copilot for Service integrates with CRM software for customer service use cases, while the Copilot in Dynamics 365 Guides is designed to summarize useful information for frontline workers in a variety of industries.
We also discuss the new update from Notion, a note-taking app enhanced with AI. Its new 2.35 update introduces a new feature called Q&A, which is currently in beta testing. This feature responds to your notes in the form of answers, making it easier to manage tasks, timelines, and priorities. (Or does it?)
News links
We discuss: OpenAI's Dev Day, 128k tokens, Claude, model distillation, ensemble models, GPT4, inference, attention mechanism, tokenizer, fine tuning, retrieval, code interpreter, JSON validator, GPT store, App store, GPT-4 Vision, ChatGPT-4 Turbo, input tokens, output tokens, prompts, context window, context side,
This is the first episode of Using AI Series 2!
I’m Alex Denne, joined once more by ML Research Scientist Alex Pap and AI Startup Founder Nitish Mutha.
We are only a few weeks ago from the anniversary of ChatGPT being released, and OpenAI have just hosted their Dev Day, with some big announcements. We’ve been busy too - attending events in the run up to the UK’s AI Safety Summit.
News links and info
Watch USING AI on youtube
This is the last episode of Using AI Series 1!
The Business of Law usually tops the list of industries to be disrupted by ML and AI innovation. Let’s get into it.
I’m Alex Denne, joined once more by ML Research Scientist Alex and AI Startup Founder Rafie Faruq.
Tell us what you want us to cover in the next series of episodes on Using AI.
On Linkedin: https://www.linkedin.com/feed/update/urn:li:activity:7077624981232177152
On Twitter: https://twitter.com/GenieAI/status/1671858707334873094
Watch USING AI on youtube (and see our AI background images) https://www.youtube.com/channel/UCHsQu4IipA7Ri2AqKcQZ1Yw
News links
Theme links
Covering areas such as Legal Transactions, Disputes, Negotiation, Litigation.
Contract drafting (because of personalisation vs templates and time and cost vs law firms) -
Document Review -
Negotiation
What our Legal Assistant does - and when it’ll be released
Crazy things from this week
Topics touched on
Copyright, the law, legal interns, legal assistants, lawyer, law firm, M&A, disputes, litigtation, paralegals, Anthropic, Claude, GPT-4, 32K tokens, 100k token limit, text generation, heuristics, citibank, citi, Microsoft Word Integrations, API calls, NDAs, License agreements, Supervised vs Unsupervised learning, in-house lawyers, OpenAI, Softbank, Venture Capital, Investment, AI Hype
That’s it for this season!
As mentioned - Tell us what you want us to cover in the next series of episodes on Using AI.
On Linkedin: https://www.linkedin.com/feed/update/urn:li:activity:7077624981232177152
On Twitter: https://twitter.com/GenieAI/status/1671858707334873094
Will AI politicians and AI political parties be a thing? Plus - AI church services in Germany, a 4-week-old startup mistral.ai raises $113m, the UK Gov <3 Deepmind, Yann LeCun (Godfather #3) tries to allay AI fears, OpenAI releases large API updates
Your host is AI geek Alex Denne, joined by ML Research scientist Alex and AI Founder Rafie Faruq.
Watch USING AI on youtube (and see our AI background images) https://www.youtube.com/channel/UCHsQu4IipA7Ri2AqKcQZ1Yw
News links
Note: the Context length for GPT-3.5 turbo has been increased
Theme links
Crazy things from this week
Topics we touched on
Stay tuned for our next episode on - AI in the legal industry
Can anyone code if they are pair programming with AI? Are these tools accurate and helpful? (Jump to 18 mins in). Join ML & Python Engineer Tom Wright, AI Startup Founder Rafie Faruq and Host Alex Denne for a discussion covering the AI news from the week, AI software development assistants and the results of the largest Turing Test ever.
Watch USING AI on youtube (and see our AI-generated background images) https://www.youtube.com/channel/UCHsQu4IipA7Ri2AqKcQZ1Yw
News links
Theme discussion and links (from the 18min mark)
Crazy things this week
Topics touched on
Unit testing, Code comments, SQL, Python, Regex, User stories, D3.js, Hans Rosling, Copilot, Codex, GPT 3.5 Turbo, Code generation, Vector diagrams, ChatGPT Plugins, Textual representation, Product management, Generative AI, Behaviour Scenarios, Tests, Acceptance Criteria, Fine tuning, IDEs, Documentation, Utility functions, Task decomposition, VSCode, API, Debugging, coding, software development, the last mile, mathematical proofs, Wolfram Alpha, AI Pair programming
Stay tuned for our next episode on - AI political parties
Next week - AI Political Parties
Let’s get cosy with our AI companions + news from this week. Join real human companions Alex Papadopoulos Korfiatis (ML Research Scientist), Rafie Faruq (AI Startup Founder) and your host Alex Denne. Yikes! Over an hour - we had a lot to talk about this week eh?
Watch this episode on youtube (and see our AI background images) https://www.youtube.com/channel/UCHsQu4IipA7Ri2AqKcQZ1Yw
News links
Theme links
Crazy things from this week
Topics we touched on
personal ai companion, roleplay ai chatbots, talking to AI, ai companionship, nvidia, chai, ChatGPT, replika, character.ai, deepmind, virtual companions, virtual friends, virtual characters, Friedrich Nietzsche, Plato, Music Recommendations, Ted Lasso, Mae
Stay tuned for our next episode on - AI and Coding
MedPaLM 2 outperforms human doctors in 86.5% of cases Meta AI’s 1.2m Token model Megabyte Meta’s LIMA: Less Is More for Alignment Photoshop & Adobe Firefly gets even better Crazy things Job interview real-time live transcription and responses Drag GAN - Modifying images by dragging certain points Prompt tips - Knowing how the model was trained to choose your prompt engineering style (e.g. if it was trained on Q&As, then format your prompt with a Q&A). - Give the model many examples of what you expect (e.g. if you want your answer in a certain style) - Specificity vs generality. The more specific you are, the better the answer you’ll get. - Always remember that the (transformer) model is predicting the next most probably sequence of tokens or words, whenever you are creating a prompt.
Links to other companies / models / startups mentioned
Clearword meeting notes summariser
Tactiq Mentions: Model conditioning, Tokens, Prompt Engineering, Model Training, Sequence2Sequence, RLHF, model weights, hyper parameters, supervised learning, unsupervised learning, annotated data, caselaw, finetuning, model selection, LLMs, Google, Rainbow Corn, Generative AI, Emergent capabilities Book: Michio Kaku: Physics of the future (misquoted as Science of the future)
The implications of Sam Altman's testimony plus prompt hacking / prompt engineering tips. EU AI Act, Open Source OpenAI Models and much more.
Links
AI News this week Sam Altman’s abridged testimony (Twitter):
Senate sub-comittee hearing in full:
Sam Altman Testimony on TikTok
Mentions:
Crazy things this week
https://mealpractice.com/generate - Effortless meal planning with AI-generated recipes
Elon saying OpenAI wouldn’t exist without him.
Discussion Points
Prompt hacking tips
Be succint and specific!
Prime the language model by providing a persona for it.
Asking the model to think through its response step-by-step “Let’s think step-by-step”
Providing clear context: Offer concise background information for guidance. Use Cases: Definition explanations, historical event summaries, concept descriptions.
Specifying the output format: Indicate desired answer structure explicitly. Use Cases: Generating lists, step-by-step instructions, summarizing long texts.
Using explicit instructions: Request specific detail or critical thinking. Use Cases: Debating pros and cons, analyzing biases, evaluating arguments.
Redundancy and rephrasing: Reinforce information by reiterating questions. Use Cases: Clarifying ambiguous topics, extracting specific details, verifying facts.
Temperature and token settings: Adjust randomness and output length. Use Cases: Creative writing, focused summaries, generating multiple response variations.
Iterative refinement: Refine prompts based on previous responses. Use Cases: Troubleshooting, problem-solving, narrowing down complex topics.
Prompt engineering: Craft effective prompts using templates or examples. Use Cases: Analogies, translating complex topics into simple explanations, generating structured responses.
Advanced techniques like asking the language model to think step by step
Shout out to DeepLearning.AI’s free prompt engineering course: https://www.deeplearning.ai/short-courses/chatgpt-prompt-engineering-for-developers/
Links to us on other podcast platforms
Spotify - https://open.spotify.com/show/5Lbw5IlX0z22rqzQheZ5My
Apple Podcasts - https://podcasts.apple.com/us/podcast/using-ai/id1683562982
Google Podcasts - https://podcasts.google.com/feed/aHR0cHM6Ly9hbmNob3IuZm0vcy9kZjg0ODk2Yy9wb2RjYXN0L3Jzcw
Amazon Music - https://music.amazon.co.uk/podcasts/d68cc64d-9854-43b2-b72c-c54486f0cb06/using-ai
Castbox - https://castbox.fm/channel/id5409720
Overcast - https://overcast.fm/itunes1683562982/using-ai
Pocket Casts - https://pca.st/8q41hgrb
Radio Public - https://radiopublic.com/using-ai-GMvYqZ
Stitcher - https://www.stitcher.com/show/1066264
RSS - https://anchor.fm/s/df84896c/podcast/rss
News from the week
PaLM2 - better at reasoning, maths, and logic, and in many cases, and does better than OpenAI’s GPT-4.
PaLM2 PDF technical report
Apps mentioned
gamma.app - Generation of the entire presentation
canva.com - AI throughout all the tools
Entrepreneur First (EF) - https://www.joinef.com/
Crazy Things
Palantir showcases “AIP” platform using LLMs in military situations 😬 (https://www.youtube.com/watch?v=XEM5qz__HOU)
Guardian article calling LLMs the largest heist in human history
The Scientist and the A.I.-Assisted, Remote-Control Killing Machine (NY Times)
Books mentioned
Innovator’s Dilemma - https://en.wikipedia.org/wiki/The_Innovator's_Dilemma
New Silk Roads - https://en.wikipedia.org/wiki/The_Silk_Roads
Groups of key words people use to actively look for AI tools with
Mentions
MedPalm2
The UK’s ICO, GDPR and Automated Decision Making (https://ico.org.uk/for-organisations/guide-to-data-protection/guide-to-the-general-data-protection-regulation-gdpr/individual-rights/rights-related-to-automated-decision-making-including-profiling/)