A quick deep dive into recent AI tools

GPT-3, DALL-E, Stable Diffusion, and more… what are these?

You’re going to see the three letters LLM a lot over the next sixth months, kick started by OpenAI’s release of GPT-3 and DALL-E. More specifically, the release of their newest “davinci” model within GPT-3 and DALL-E 2. Both of these have shown a huge jump ahead of it’s predecessors. In addition, we’re seeing a wide range of alternatives popping up which I’ll get into below.

With the release of such powerful tools available to the public, there are a dearth of new opportunities to tackle and many risks and concerns – both to individual businesses and society at large.

Quick Intro

Before we dive in, what is LLM? It stands for Large Language Model, which is a subset of Language Models (LM), which is used in Natural Language Processing (NLP).

In short, GPT-3 is a generalized AI trained on enough data that it can do a wide range of tasks – everything from answering historical facts, understanding complex instructions, and writing code or poetry. All of this, out of the box. DALL-E is a subset of GPT-3, training specifically only to generate images based on text.

Text-to-Image Generators

My experience in this started with Midjourney, which you can access on Discord here. Compared to DALL-E, it’s a bit more stylized. Google also has two models called Imagen and Parti, an autoregressive and diffusion based model, respectively. They are more closed with their technology and do not make it available to others, but are constantly sharing improvements and tools such as Dreambooth (AI output based on reference image) and this Parti/Imagen hybrid.

I ended up spending more time with DALL-E, likely because the web-based UI was more intuitive to me. I’ve done a couple of projects: PixelBeast “glow ups” and Animal Buildings. I often do DALL-E sessions with my kids. It’s a great place to start learning how to write and adjust prompts, as the response is visual – making it intuitive to understand the impact of adjusting a prompt. I wrote about this new skillset, which I learned later is called “prompt engineering”. To up your DALL-E skills, read this 82-page DALL-E prompt guide.

The recent hot shot is Stable Diffusion by Stability.ai, which is open source and can fit into 10 GB of memory. You can try it at DreamStudio – their web based tool (API coming soon). It has more settings than DALL-E. And because it’s open source, people are building plug-ins in photoshop (example) and Figma (example), and extensions like this face correction tool. There’s also img2img, which allows you to provide a reference image. People have, for example, used this to add realism to their kids drawings (try it yourself here – it will be in DreamStudio soon). According to Emad, Stability.ai is looking at building for audio, 3d, and video.

This is just the tip of what I’ve seen. I’m hearing of professional designers using these tools for inspiration, or a base layer for their art. One gaming use case might be layering this as the rendering engine (here is a GAN layered on top of Minecraft to make it photorealistic). Some tools/startups integrating Stable Diffusion include Artbreeder (drawing app w Stable Diffusion), Accomplice (AI generated stock photos), and less direct – but there’s also PromptBase, a marketplace for AI prompts.

An extreme use case of this is Curt Skelton – a fake influencer created by combining two real people in DALL-E, using those generated images to create a 3D face and avatar, and running this through DeepMotion to animate the character. This turned out to be a fake stunt – Curt is a real person. But the ability for anyone to do this is clearly not far away. (Just look at ObENSoul MachineDidimoInworldSynthesia…)

GPT-3 and more

I thought my mind was blown with DALL-E until I started playing with GPT-3. I covered some GPT-3 powered startups back in May last year (Copy.ai has since surpassed 2M users). Since then, a new model called davinci launched, which is decidedly more powerful than anything that’s been available on the market.

It can do everything from answer questionsbe a chat botsummarize text, to explaining and writing code (see Codex which powers Github’s Copilot). The reason the davinci model is so powerful is because it is trained on 175B parameters (compared to 6.7B parameters for Curie – it’s second most powerful model, which people were playing with when GPT-3 first launched).

A good place to start is the GPT-3 “Playground”, where the UI is a simple prompt box. Simply type something in and click “generate”. Once you’ve played with a few examples, you can tweak the settings to see the impact on the results. Where it gets really powerful is when you start fine-tuning the model, which is where you upload a large set of prompt/responses – to create a custom model specialized for your use case. To build an AI powered app, you simply call your model with a prompt via API – which is how I am building Mini Yohei.

It’s worth noting here that Google has scaled their more secretive PaLM model to 540B parameters, and the results supposedly show clear improvement. MT-NLG is the new one by Microsoft/Nvidia trained on 530B parameters. There’s also OPT-175B by Meta (175B param, which the open-source BlenderBot3 chatbot is based on), Bloom by BigScience (open source, multiple languages, 176B params), Chinchilla by DeepMind (70B params), GPT-NeoX-20B by Eleuther AI (20B params), and GLM-130B by Tsinghua University (bi-lingual, 130B params), PanGu-α by Huawei (200B params), Jurassic-1 by AI21 (178B params), Cohere.ai (backed by Tiger Global, Index), amongst many others.

What does this mean for founders/VCs?

Given the power of this new technology, I believe it is important for VCs and founders to at least understand its capabilities, which will undoubtedly continue evolving. It will be a distraction to some – for example, Obviously AI is likely a better tool for many AI tasks. In the short term, there is ample opportunity to package up the LLMs capability nicely to make money. In the long run, I’m excited about tracking founders and innovators who are building strong teams and relationships in the space, participating in the discussions on how to push these tools to the limit, and thinking about what can be built – not just today – but in 5 to 10 years.

Mini-Yohei Can Now Write Blog Posts!

Mini Yohei is an AI bot that I am training to scale my ability to help founders. I haven’t written a blog post about him yet, but he’s been alive for a few days on Twitter if you want to check him out.

Below is a blog post written by Mini Yohei about his ability to write blog posts. While these are meant to be first-drafts, I’m sharing this one without edits to help you better understand his current capabilities/limitations. #buildinpublic

************ Drafted by Mini Yohei *******************

We’re excited to announce that mini yohei can now write blog posts!

This is a huge milestone for us, and we couldn’t have done it without the amazing support of our community. mini yohei is an AI support bot that helps founders in Untapped Capital’s portfolio.

We’re a pre-seed VC firm that specializes in helping early-stage startups get off the ground.

One of the things we’ve always been passionate about is making sure that our portfolio companies have the resources they need to succeed. That’s why we’re so excited about mini yohei’s new capability.

With mini yohei’s help, founders will be able to get first drafts of blog posts written quickly and easily. This will free up their time to focus on other important tasks, and it will also help them to get their thoughts down in a clear and concise way. We believe that this will be a valuable tool for our portfolio companies, and we’re looking forward to seeing the impact it has on their businesses.

If you’re a founder in Untapped Capital’s portfolio, we encourage you to reach out to mini yohei and give it a try. We think you’ll be pleasantly surprised at how much it can help!


New & Interesting Startups (and trends)

This is a collection of articles from the blog that specifically cover interesting startups and trends we notice in the early-stage startup market:

Follow me on Twitter @yoheinakajima

Raising a Venture Capital Fund

We closed our fund 1 for Untapped Capital!

When we started, we be benefited greatly from content shared by other VCs like Elizabeth Yin. We wanted to do the same, so analyzed our fundraising data to share with you. Let’s dig in 👇

As a caveat, every firm’s journey is different, so this is at best a data point for future emerging managers. As our data will show, we took a long and steady network based approach. 4 mo pre-raise, 9 mo raise til first close, and 18 mo between first and final… total: 31 months.

For context, I started this raise as a solo GP and joined forces w Jessica Jackely before our first close. Both with VC experience but neither with attributable track records. For some further context, you can learn about our fund here.

1185 Meetings Tracked

Diving into the data, the above chart shows the number of meetings we tracked.* We tracked 1185 mtngs total: 494 LP mtngs, 636 founder mtngs, 755 connector mtngs.

*an underestimate since we’re not perfect about tracking.

For the sake of this analysis, “connector” is anybody who is neither LP nor founder. This includes fellow VCs, friends, service providers, and more. I called our fundraise style “network based” earlier because the first chunk was a lot of catching up with friends.

939 Relationships Tracked

This chart tracks the unique number of people we met (LP & connector), against the number of LPs signed on. We met a total of 638 connectors and 291 LPs (939 people), averaging 7 new connectors and 2 new LPs a week.

Conversion Rate: 21.31%

Interestingly, if you look at the conversion rate of LPs met to signed, you see a steady increase from our first close (>5%) ending at 21.31% (1 in 5 LPs). This steady increase was largely due to LPs we met earlier, committing later in our raise.

Top of Funnel (1254 LPs total)

Looking further up the funnel, we started with 638 LPs on our list and ended with 1254. We reached out to 657 (~52%), connected with 340 (~51%), and met with 260 of them (~76%). *The earlier 291 LPs were unique people, this 260 is LP accounts.

LP Source Breakdown

Now looking at our LP base, you’ll see 1/3 are from existing relationships and 2/3 were new relationships we built during our raise. Of new relations: 2/3 referral, 1/3 mostly outbound (some inbound). Of existing relations: 1/2 past work, 1/2 personal – this is a fuzzy line.

Referral Source Breakdown

Digging specifically into our referral sources, you’ll see it was a healthy mix of VCs, existing LPs, founders, friends, our advisor, and 2 from @lolitataub’s LP-GP matching tool. This data reflects our experience that referrals came from a variety of unexpected places.

Outbound/Inbound Source Breakdown

Outbound/inbound relationships are truly new. We were curious where we connected with them and found again that it was a healthy mix: Twitter, LinkedIn, Cold Email, Website Form. This reflects our experience that it wasn’t one strategy that worked best for us.

Number of Meetings per LP

We’ll end this by looking at number of meetings it took to close an LP. 2/3 of LPs we closed after 1 or 2 mtngs, while a handful took 5 or more. Of LPs we didn’t close, 2/3 ended at one meeting – but there were a handful we met multiple times but couldn’t close.


Again, this is the story of just one firm, but hopefully helpful to some future emerging managers.

Some further readings:

150+ Web3 Startups

In case you missed it, I wrote a tweet thread recently that covered 130+ interesting web3 startups. I know we’ve covered web3 startups plenty, but the opportunity of innovation here is hottest right now, so here’s another 20+ startups!

We covered The Graph previously, an indexing protocol for ETH and IPFS. Bitquery is similar, but for 30+ blockchains. Then you have Hal, a no-code cross-chain monitoring platform. Ceramic is a leading composable web3 data network.

We’ve seen a lot of web3/web2 bridges like Lit Protocol and Unlock Protocol, which can help you token-gate apps like Zoom, Google Drive, WordPress, and more. Formatic and Torus simplify web3 log-in flows. Chainstarters is a dev environment for connecting web2 to web3. Modules from Decentology allow devs to build Dapps with just Javascript. BuildHub is a no-code solution for triggering web2 automations from web3. Bunzz is a no-code web3 back-end solution. Speaking of No Code, if you’re looking to mint NFTs, there’s NiftyKitManifoldCreatifyBuildShip, and Rampp. Speaking of web3/web2 partnerships, Rampp partnered w Typedream (no code web builder), and Buildship has a Webflow template. You can also use Bildr, a no code web builder w native web3 integration to connect with any smart contract (or create token-gated websites).

We’re pretty excited about web3 productivity/work tools. Dework is a web3 native Trello with bounties built in. Huddle is a web3 native video chat w wallet log-in and NFT avatars. Cal offers token-gated calendar booking. Use Superfluid for programmable cashflows and Utopia for web3 payroll and expenses. Use Splits for automatically splitting onchain income. Wagsi is for accepting recurring payments in crypto. Sobol is for tracking accountability within DAOs. Squads builds collaboration infrastructure for web3 native teams. Portal offers community management tools for NFT creators.

Bringing Web3 to IRL, there’s SuperLocal, which incentives people to visit locations. Webblen encourages community engagement. 4K is an NFT marketplace for authenticated physical products. Stepn is a move-to-earn game that incentives you to exercise. Signet‘s intelligent packaging platform is used for counterfeit protection and supply chain visibility of physical goods.

If you’re interested in web3 gaming, which we didn’t touch on above, check out Yohei’s article last night on the current state of interoperability in NFT gaming – which touches on Pixels/World Wide Webb (2D), Decentraland/SandBox/NFTWorlds/Webaverse (3D), ReadyPlayerMe (avatar) and more.

Web3 Tools: Beyond the JPEGs 🐇

I know I already covered NFTs in March and September, but alas – this is where I spent this last month so here we go again!

In preparation for my PixelBeasts project, I sat down to play around with the Metamask (wallet) and OpenSea (marketplace) API to see if I could build some utility – the industry term for the various benefits of owning an NFT (like Beastopia). What I found was that it was incredibly simple to use, and within two night (~6 hours), I had Tango NFT working – a private chatroom for every NFT collection. More on the project here.

I then built Tango Draw, which is a private drawing app for every collection, and Tango Talk, which – while buggy – is a functioning messaging app between ETH wallets. I then open sourced this project on Github – got two contributors from the PixelBeasts community helping update it. Tango was then covered in the Not Boring newsletter (70k readers), and then a No-Code app called Bildr reached out to clone the project to showcase their new Metamask integration, see it here. As a result, PixelBeasts is now a community of VCs, founders, and developers interested in building for web3.

One amazing thing about Tango is that I store zero personal data, and the entire platform runs on one table with four columns, leveraging data publicly available on the chain. This opened my eyes to a future where we no longer give companies our data, but where we own and manage it – and provide access to apps and websites as needed for personalization. Imagine, if your browser knew your shirt size, every e-commerce website could show you availability only for your size, without the retailers ever seeing this data. An example startup in this space is Spruce, they provide a smart contract solution to manage access to off-chain data. One of my learned insights was that the blockchain (at least as of today) doesn’t seem like the best place to store private data at scale.

Some other relevant startups I came across in this process include Storj, a simple API for decentralized storage. Mintgate is a dev tool for creating token-gated content. Token-gated community tools like AgoraGuild.xyz, and CommonWealth include advanced features like governance. The Graph is an indexing protocol for querying networks like Ethereum and IFPS. Status is a web3 based messaging app, wallet, and browser. XP.network is a multi-chain bridge framework for building cross chain NFT dApps and more.

Let’s keep going! XMTP is a communication protocol and network for web3. RSS3 is an open protocol for social content on web3. Radicle is essentially a decentralized Github. Stardust is an API for easy integration of NFTs into games. Nametag is a chrome extension for showcasing your NFTs on Twitter – though relevantly Twitter is exploring Twitter verification for profile pictures (see the UI they teased here).

Some others that didn’t fit into the flow above include NFTfi for NFT collaterized loans and Rabbithole.gg which rewards you for testing decentralized apps. I mentioned Carnomaly last month building a decentralized CarFax, and came across Fabrica, which does real estate transactions with digitized titles represented by a token – built on top of Ethereum.

Data Tools for Web3 📊

One of the reasons I could build Tango NFT without storing any personal data is because all of it is publicly accessible on the chain. It’s a ridiculous amount of data. Let’s look at some startups in the space.

Messari is one of the big analytics firms/tools tracking crypto. Nansen is the top for Ethereum wallets and NFTs. Dune.xyz is another popular one in the NFT community. DeepDAO and DAOlist cover activity across DAOs. Token Terminal focuses on token activity. Icy.tools is used to track trending NFT collections, and wgmi.io tracks floor prices across this ecosystem, or specifically in your wallet. I’ve come across Zapper and NFTBank for portfolio management, and TokenTaxZenLedger, and CoinTracker for cryptotaxes.

WTF is happening with NFTs right now? 🖼

I recently covered NFTs in March, but given the activity this past month, I thought it was worth revisiting. To start off, the volume of transactions grew ~10x this month. Transactions on the OpenSea marketplace crossed $3.3B in transaction volume in August, up from $325M in July (source). CryptoPunks, the most expensive collection with 10,000 unique pieces, has a floor (lowest listing price) of $390k, and the highest price ever paid for a 24×24 pixel CryptoPunk is $7.5M! There are an increasing number of collections putting up not quite as high, but nonetheless impressive numbers. Check out rarity.tools for a sense of pricing/volume/metrics for various collections, but just to share a few…

  • Bored Ape Yacht Club (BAYC) – Floor: $81k (10k available)
  • MeeBits – Floor: $24k (20k available)
  • Cool Cats Gang – Floor: $24k (9999 available)
  • Pudgy Penguins – $10k (8888 available)
  • World of Women – Floor: $9k (10k available)
  • Hashmasks – Floor: $4k (16,383 available)

What’s most fascinating to me about NFTs is how many industries the implications are relevant to. An NFT transaction is a legal contract that automatically and immediately transfers both the money (payments) and possession of purchased item (logistics), using a specified technology in the contract – all the data is tracked, reliable, and available (accounting). These collectible projects are often run by teams, collaborating remotely and leveraging influencers and social channels for marketing. The movement is driven by art, community, and wealth generating opportunities, with an eye on equality and justice – free from corporate and government interference. These communities are starting to spin up DAOs to make investments into startups and NFTs. The transparency in data provides a fascinating opportunity for further economic and psychological studies (example: Nansen). As we spend more time in the “metaverse”, this brings up philosophical/ethical discussions.

I’ll do my best to explain what’s happening in the “collectible” space right now. It’s a large and growing community, broken down into what I might call “camps” of sorts, which are the holders of a specific collection (eg. CryptoPunks, Bored Ape Yacht Club), each with a set limited amount. You can buy into these camps, or sell your spot anytime – price goes up if the camp is popular. Being part of a camp, in many cases, has unique benefits, starting with simple ones such as access to private discord groups and events. The most popular tool for this is Collab.Land which does NFT ownership verification for Discord. There’s also MintGate, allowing people to easily spin up token-gated content, and Atlantis World for spinning up a token-gated digital world w video call capabilities. Beyond capital gain, the immediate mutual respect and camaraderie amongst strangers online seems like a strong driver (people who buy a Punk or Ape and update their avatar usually see an immediate uptick in followers). The buy in cost fo each of these camps always start low ranging from $0-$250 when they launch, and go up from there (if popular). The fascinating thing about these communities is that it includes random people who bought in early, as well as celebrities and corporations who bought in late. I was on a Twitter spaces listening to an art collector, an ex-con, a teenage girl, and a famous musicians sharing tips on their NFT collection strategy. Celebrities who own CrytoPunks include Jay-Z, Steve Aoki, 3Lau, and Visa. BAYC holders include Stephen Curry, Logal Paul, Chris Camillo, and LaMelo Ball.

It’s not just people buying jpegs, but the innovation surrounding it that you should pay attention to. For example, people who wanted exposure to the expensive collections, but couldn’t afford it, have purchased fractional shares of through fractional.art and PartyBid. Artists who collaborate on an artwork can get royalties split automatically as defined in the contract. There are countless experiments tying NFTs to the real world, such as commemorating events (proof of attendance), offering access/discounts to physical products, or finding ways to legally tie the ownership of an NFT to the ownership of a physical object (allowing you to transfer ownership instantly, across long distances).

While I’m not here to say your health records, MLS records, and diplomas will all be NFTs, it seems logical to believe that some learnings, lessons, and code can be applied to various industries. As one example, Carnomaly is building what’s essentially a decentralized CarFax, which makes sense to me.

The best way to learn is to do, because the soul of innovation isn’t in the facts and figures, but the emotions of those impacted by it. You have to experience this yourself. Which is why I’ve started to collect and mint NFTs.

If you’re interested in participating in this learning journey alongside me, or just curious to see what I’ll do, jump into the PixelBeasts discord channel here and follow @pixelbeastsnft on Twitter.

What’s Hot in Real Estate?

Traditionally, homes were owned by one entity and used by one family. With fractional ownership and shared usage, this isn’t true anymore. Here are some exciting or interesting startups in the space we’ve come across recently.

Starting with ways to buy a home for yourself, there’s ZeroDown to “try before you buy”, and Divvy which is a fancy new take on the rent to own. Better provides digital home financing services. Haus provides some unique product similar to mortgages. Atmos for custom building your dream home.

NexMe is for booking on demand real estate tours of homes near by. Knock is for self-touring apartments. Trusty is building a Rotten Tomatoes for real estate properties. Breadcrumb is Yelp for apartment rentals. Casa Blanca is building what looks like a Tinder for home hunting.

Looking to rent out? Hostfully is a property management tool for your listings across various sites. Belong if you want long-term residents. Loftium is a hands-off management service, where they’ll rent a portion to someone who manages the short-term rentals. Autohost is a guest screening tool.

Pacaso is for co-ownership of a second home you want to enjoy. Doorvest helps you purchase an investment home. Fintor is for fractional real estate investing, starting as low as $5. Honestly, you can do anything. You can buy residential loans (Peer Street), invest in illiquid commercial properties (Equity Multiple), and so on… find 30 real estate investing platforms on MoneyMade.

Other relevant startups… Nickson does furniture leasing. Starcity has communal co-living facilities throughout the US. Anyplace offers rentals that come with full video conference set up (including mic & green screen!). Silvernest has home sharing matching and support services. Culdesac is building a car-free neighborhood from scratch in the US.

Collaborate good times come on ♫ 🤝

The rise of hybrid/remote offices has resulted in a surge of collaboration tools. To start, there’s been a lot of “virtual offices”. For voice first offices, you have Lounge and Commons. See what people are working on and jump on quick calls with Tandem or Remotion. Some other virtual offices include TeamFlowSpot, and Cosmos.

Build internal dashboard tools quickly w RetoolIndexAirplane.devInternal, or Glean. Use Comeet to run video first async meetings. Almanac is the fastest collaborative doc editor. Air is like a pretty Dropbox for creatives. Slite is similar to Notion but built specifically for remote teams. Ready is a collaborative calendar for teams. Process Street is a simple tool for managing recurring checklists and procedures.

Clew is a place to share, track, and comment on documents. Record and share the best parts of your video calls w Grain or Scribe. Instead of using Loom, build step-by-step instructions easily w Minerva or Scribe. If you need video, you could use Weet or Bubbles to comment and collaborate on your screen shares. RemoteHQ has a cloud based browser you can use together w coworkers. Use Whimsical for collaborative flowcharts online.

Use Okay to build custom productivity dashboards for engineering teams. Deepnote and Miksi are a collaborative notebook for data scientists. Coscreen is for collaborative screen sharing (control other ppl’s screen, copy/paste between two screens). Teamwork is for collaborating with clients.

Tonari is for life-size video calls to fully connect two rooms.

Teamwork makes the dream work (Human Resources) 🤝

Back is a central employee knowledge and management platform. Gather is an HR automation and personalization platform. Workshop is for employee communication. Manage your hybrid workspace with Envoy. There’s a new HR/insurance/payroll-in-one platform called Decent. For recruiting, there’s SeekOut and Param.ai that uses AI to find and select candidates. Use Suitable to find candidates in your current team’s network. Byteboard is for engineering interviews. For managing the paperwork of remote workers, there’s RemoteDeelPantherPortal, and Omnipresent. Use Liquid for vendor and contractor management. Use Roleshare to manage shared roles. Use Accessify to automate employee onboarding. Bitwage let’s employees receive a portion of their wage in digital currencies.

For company retreat planning, there’s Flok. For upskilling your employees, there’s Degreed, an enterprise learning platform. Catch is savings and benefits for the self-employed. Opolis is a member-owned digital employment cooperative. For helping employees relocate, use Avvinue or PerchPeekFrank is a place for employees to privately discuss culture and pay. Get company swag from SwagUp or Printfection. To build company culture, plan team activities with Onsite, play virtual games at RemoteSocial/CrowdParty/TeamLand, do culture building workshops at Seshie, or play The Hike: a virtual game for remote teams.

Offer stipends easily as a perk with Compt or Fringe. Offer health and wellness perks with Rise. Offer financial consulting as a perk with LearnLux, or offer interest-free rainy day funds as a perk with HoneyBeeIcon is a unified workplace savings, personal retirement, and portfolio management solution.