Really feels like there's been a phase change in AI x Crypto. More cooperation between projects, standards forming etc. It feels like we have all of the building blocks in place. Time to put them together...
Follow
Richard Blythman
@richardblythman
Co-founder | Decentralized AI | Regenerative AI | Machine Learning Engineer | Web3 Developer | Fluid Dynamicist
Richard Blythman’s Tweets
25/ Human Controlled Outcomes: Automation needs to be human led. Machines abilities to make decisions on their own will continue to improve, but it's crucial to keep people in control of outcomes.
1
Show this thread
24/ Human in the loop: Humans provide data to help guide the IDW when it gets stuck. Flows that incorporate HITL reach out to humans on different channels to get the needed information. Updates their knowledge bases & skills. Humans can feed data or script interactions.
1
1
Show this thread
23/ Anomaly Detection: A Process Mining pattern for detecting anomalies in data, such as events or deviations from what is standard or expected. Can also be used to detect anomalous absence of data.
1
Show this thread
22/ Process Mining: Analyze data with the aim of identifying patterns, inefficiencies and opportunities in both current and historical processed and events.
1
Show this thread
21/ Predict: Use past interactions and contextual data to predict what a user might be trying to do, often suggesting possible outcomes. Aims to eliminate unnecessary steps to make the conversation as efficient as possible.
1
Show this thread
20/ Teach: Show users how to do something, often launching from Q&A. Provide a series of lessons and instructions, which can occur in a single session or across multiple sessions.
1
Show this thread
19/ Share: Send information to people who need it. Helps disseminate information in a contextual way. Flows that use Share send out messages or links bearing useful information.
1
1
Show this thread
18/ Coordinate: Get several participants working together around a particular goal. Used to e.g. schedule a meeting or gather shared input. Complex pattern that uses flows such as Chase, Track and Transact working together.
1
1
Show this thread
17/ Promises & Assignments: Someone promises to take care of a certain task or assigns someone else to it, for the next time they log in. Takes the form of a queue of assignments that is revealed when someone makes inbound contact. Has an asynchronous component.
1
Show this thread
16/ Negotiate: Build negotiation into the process with IDWs. Users can gently persuade the IDW to bend the rules their way e.g. asking if they can check in to a hotel room early
1
Show this thread
15/ Transact: Help users to accomplish a particular task w/ a goal in mind and a desired outcome e.g. scheduling an appointment, ordering a product, making changes to a setting, adding new communication preference. Has a structured script that needs information to complete task.
1
Show this thread
14/ Concierge: Greets a user and evaluates their needs. Uses patterns such as Contextualize and Q&A. Then uses Guide to connect users to other skills in the ecosystem that will help them achieve their goals.
1
1
Show this thread
13/ Guide: Bring user from point A to point B e.g. checking in each day to ensure users stay on track with a particular goal. Keep in mind progression in sequence with the milestones or ultimate outcome they're meant to achieve. Critical for Concierge skill.
1
1
Show this thread
12/ Contextualize: Machine queries contextual storage to continue the conversation from that context e.g. time of day, location, task at hand or prior conversations. Cuts out the need for starting from the beginning, asking repeat questions etc. Goes hand in hand with memorize.
1
2
Show this thread
11/ Callback: Focused on resuming a prior activity. Geared towards pausing an activity and setting a follow up, depending on time elapsing or the emergence of a new piece of data.
1
Show this thread
10/ Track: Similar to memorize. Track how many times a user goes from point A to point B. Flows that track are keeping in mind a current state, as well as prior states.
1
Show this thread
9/ Remind: Proactive pattern that gives users information at a particular time and in a specific way, in order to take action. Sent out over whichever channel the user prefers.
1
Show this thread
8/ Memorize: Used to understand the common conversations & questions that users engage in, preferred channels etc. Stores information to be used for reporting, making enhancements to the knowledge base & providing future context. Helps to build relationships with users.
1
1
Show this thread
7/ Drip: A series of announcements that can be used for reporting and making announcements that don't require immediate feedback. Often represents a content journey e.g. spaced out messages that go out to first-time customers as part of an onboarding experience.
1
Show this thread
6/ Nudge: A soft push towards a desired outcome in a manner that is less intrusive than Chase. Designed to provide extra information in a structured way to motivate users to take a particular action.
1
Show this thread
5/ Chase: Like reminders but more aggressive. This flow activates continuously until certain criteria are met e.g. repeat a query with a user until an answer is obtained, or try with different users. Often successful resolution will require escalation.
1
1
Show this thread
4/ Find: Differs in that the machine queries external sources like the internet via API and displays the results to the user. This flow should be employed of Q&A fails.
1
Show this thread
3/ Q&A: One of the most basic flows. User asks a question and the answer is provided from the knowledge base with a source. This flow should turn to Human in the Loop if it doesn't have an answer.
1
1
Show this thread
2/ Terminology: An intelligent digital worker (IDW) is made up of skills, composed of microservices. Flows are sequences of skills used to perform tasks. Good conversational AI should use many of these flows in tandem. Some examples of flows:
1
Show this thread
1/ Key design patterns for conversational AI from the book Age of Invisible Machines, A Practical Guide to Creating a Hyperautomated Ecosystem of Intelligent Workers by Robb Wilson
3
1
15
Show this thread
Has anyone thought about LLM agents in the context of finite state machines, or come across some reading material on this?
6
13
Fireside chat about the next generation of AI Platforms, AI Hubs and the AI Commons.
4
15
One fun example of different fields coming together. performing in a cube (measured with 3D sensors and BCI) with projections from AI images generated by Stable Diffusion . Art by and .
1
8
Show this thread
Pleasure to give a talk on what we think the next generation of platforms for generative AI might look like, and why we should make use of decentralized infrastructure.
3
14
Show this thread
The AI x Crypto community came together for the 1st time in Montenegro last week with appearances by & many more. Excited to see so what this group (with backgrounds from blockchain & AI developers, political scientists to artists) can build.
7
4
41
Show this thread
Video with lots of inspiration for how to approach alignment in the intelligent systems we're building e.g. "We transcend ourselves by internalizing other people's perspectives on us".
Reminds me of one of my favorite quotes: "Through others we become ourselves" Lev S. Vygotsky
Quote Tweet
I implore you to seriously consider universal enlightenment as one potential result of AGI. #AGI #AI
Full video essay here: youtube.com/watch?v=A-_RdK
2:16
29.1K views
1
1
5
Are you ready for decentralized AI summer? 👀 Join us at the very first AI x Web3 hacker house in Lisbon on 9-12 May.
5
25
Trend towards paid APIs continues (first Twitter, now Reddit). I think this might accelerate as ChatGPT + Plugins start to steal attention.
Quote Tweet
Off the cuff reactions to Reddit's pay-for-API plan:
Probably bad for:
- computer and social sciences
- CS education
- 3rd party client devs and users
Probably good for:
- Putting market value on data (helps data co-ops?)
- Seeing impact of Reddit data on AI capabilities
Show this thread
1
11
Must read thread. What type of world orders are compatible with increasingly powerful AI? So far, we're trying AI anarchy and AI authoritarianism. We need to move more towards AI liberal democracy and AI pluralism.
Quote Tweet
Worrisome paper showing something many suspected. I think beside standard AI alignment/safety there is a need to work on tools and institutions to resist AI-empowered authoritarianism. twitter.com/cameron_pfiffe…
Show this thread
1
3
8
Next generation of alignment research: Simulating an ecosystem of ChaosGPT agents together with OrderGPT agents.
2
1
13
We started laying these ideas in our NeurIPS 2022 workshop paper "Decentralized Technologies for AI Hubs"
2
4
Show this thread
And here's what a decentralized AI platform might look like, running on top of the decentralized AI stack
Quote Tweet
Nice, you just described our longterm vision for an AI Commons 
Now just need to get it funded...
Storage: @IPFS, @CeramicNetwork
Data: @OceanProtocol @Nevermined_io
Access: @LitProtocol
Compute: @BacalhauProject @GensynAI
Training: @togethercompute
Governance: @AragonProject twitter.com/vishalsachdev/…
Now just need to get it funded...
Storage: @IPFS, @CeramicNetwork
Data: @OceanProtocol @Nevermined_io
Access: @LitProtocol
Compute: @BacalhauProject @GensynAI
Training: @togethercompute
Governance: @AragonProject twitter.com/vishalsachdev/…1
5
Show this thread
The regulation: “Content generated by generative AI should embody core socialist values and must not contain any content that subverts state power, advocates the overthrow of the socialist system, incites splitting the country or undermines national unity,” Nope.
Quote Tweet
Even China recently stepped up to regulate A.I., as reported recently in FT.
Humanity can **only** win the A.I. race by being as aggressive about safeguards and guardrails for A.I. as we are about developing it.
ft.com/content/755cc5
Show this thread
1
1
7