A startup called To read has just come out of hiding with technology that helps AI models remember users and conversations. Built in the famed startup factory of UC Berkeley’s labs, it also announced $10 million in seed funding led by Felicis’ Astasia Myers, at a post-money valuation of $70 million.

Letta also advantages from the backing of some of the most outstanding angel investors in AI, including Jeff Dean of Google, Clem Delangue of Hugging Face, Cristóbal Valenzuela of Runway and Robert Nishihara of Anyscale.

Founded by Berkeley PhD students Sarah Wooders and Charles Packer, it’s a highly anticipated AI startup. That’s since it’s the brainchild of Berkeley’s Sky Computing Lab and is a business arm of the popular open-source project MemGPT.

Berkeley’s Sky Computing Lab, led by renowned professor and Databricks co-founder Ion Stoica, is the offspring of RISELab and AMPLab, which spawned corporations corresponding to Anyscale, Databricks, and SiFive. In particular, Sky Lab spawned quite a few popular open-source large language model (LLM) projects, including Gorilla LLM, vLLM, and the LLM structured language SGLang.

“A lot of projects came out of the lab very quickly, within a year. Just people sitting next to us,” Wooders described. “So it was an amazing time.”

One such project is MemGPT. It is so popular that it went viral even before its release.

“Someone beat us to it,” Packer told TechCrunch. The founders had published an information document on Thursday, Oct. 12, 2023, and planned to publish a more detailed document and code to GitHub the following Monday. Some random person found the document, posted it to Hacker News on Sunday, and “it went viral on Hacker News before we had a chance to properly publish the code or publish the document or, like, start a Twitter thread or something like that,” he said.

The reason for the excitement was that MemGPT alleviates a pernicious problem for LLM: In their native form, models like ChatGPT are stateless, meaning they don’t store historical data in long-term memory. That’s problematic for AI applications that depend on learning from and understanding a user over time—from customer support bots to apps that track health care symptoms. MemGPT manages data and memory so AI agents and chatbots can remember previous users and conversations.

The newspaper post stayed at the top of Hacker News, a preferred developer site run by Y Combinator, for 48 hours, Packer said. So he spent the weekend and the next few days answering questions on the site, attempting to get the code ready for release. Once the project was continue to exist GitHub, a link to it went viral on Hacker News, again. Interviews on YouTube and tutorials, Medium posts, 11,000 stars and 1.2k forks on GitHub happened quickly.

Myers of VC Felicis also learned about Wooders and Packer while reading about MemGPT and immediately realized the business possibilities of the technology.

“I saw that paper when it came out,” she told TechCrunch, and immediately reached out to the founders. “We had an investment theme around AI agent infrastructure, and we appreciated that a really important piece of that was managing data and memory to make these conversational chat bots and AI agents effective.”

The founders continued to virtually drive down Sand Hill Road, on Zoom calls with enterprise capitalists before selecting the one who first loved them.

In the meantime, Stoica was brokering connections with Dean, Nishihara, and other outstanding Silicon Valley angel investors. “A lot of the Berkeley professors, just by virtue of being at Berkeley, are very well-connected,” Packer recalled, describing how easy the angel investor process was. “They have their eye on projects coming out of this lab that are going to be commercialized.”

Competition and the threat from OpenAI o1

While MemGPT is already available and in use, the business version of Letta, Letta Cloud, isn’t yet open for business. As of Monday, Letta is accepting requests from beta users. It will offer a hosted agent service that enables developers to deploy and run stateful agents in the cloud, accessible via REST APIs, a programming interface that may maintain state. Letta Cloud will store the long-term data vital for this purpose. Letta can even offer developer tools for constructing AI agents.

With MemGPT, Wooders sees a wide selection of applications. “I think the most common use case we see is basically highly personalized, highly engaging chatbots,” he says. But there are also novel applications, corresponding to a “chatbot for cancer patients,” where patients upload their history after which share their current symptoms so the bot can learn and offer guidance over time.

It’s price noting that MemGPT isn’t the only one working on this. LangChain might be its best-known competitor and already offers business options. Major modelers also offer tools for creating AI agents, corresponding to the OpenAI Assistant API.

And OpenAI’s latest o1 model could make the need for state fixing a moot point for users. Because it’s a multi-stage model, it essentially needs to take care of some level of state to “think” and fact-check before responding.

But Wooders, Packer, and Myers see some key differences between what Letta offers and what the 800-pound gorilla of the market, OpenAI, does. Letta says it can work with any AI model, and it expects its users to make use of many of them: OpenAI, Anthropic, Mistral, their very own models. OpenAI’s technology currently works only with itself.

More importantly, Letta uses the open-source MemGPT code and firmly sides with the open-source side of the FOSS vs. black-box LLM debate, arguing that open-source is the more sensible choice for AI application developers.

“We’re positioning ourselves as an open alternative to OpenAI,” Packer says. “I think it’s actually very, very hard to build very good AI applications, especially when you’re after something like hallucination, if you can’t see what’s going on under the hood.”

French startup Jumpa contemporary take on the umbrella company concept in France, raised €11 million (about $12 million at current exchange rates) in a Series A funding round.

Jump offers full-time contracts to freelancers in search of the soundness and benefits of full-time work. It acts as an administrative companion only, and employees remain independent—they’ll work with multiple clients and negotiate their contracts directly.

Breega is leading today’s funding round, with Index Ventures and Raise Ventures also participating. The startup previously raised €4 million (around $4.5 million) in 2021.

Once signed up, freelancers can invoice their clients through Jump’s platform, and at the tip of the month, they’ll create pay slips and get paid. This feature alone implies that freelancers can set a pay schedule for themselves that may work all yr long—even during those slow summer months.

And with a everlasting contract, employees are registered with the national health system and may contribute to the national pension scheme. Jump also offers medical insurance contracts through Alana, food vouchers through Swile, access to worker savings schemes and more. In France, a everlasting contract can be particularly helpful when you are attempting to buy a house and are negotiating a mortgage with a bank.

There are some trade-offs, though. Corporate dues are deducted out of your salary, and Jump itself costs €99 per thirty days. But whenever you’re a freelancer, money is simply a part of the equation. I see numerous freelancers who want the very best of each freelancing and full-time work. So far, the startup has managed to persuade 2,000 freelancers to jump.

The startup also recently launched a free offer for freelancers just starting out. It features a free, skilled checking account with a virtual debit card that works with Apple Pay or Google Pay. There are also a couple of software features that may help you invoice your first clients, akin to a built-in invoicing tool and a dashboard for tracking financial performance.

“This more or less corresponds to the way freelancers work: they often start with the basic French freelance status, and then move to another status when they start to feel the limitations of their status and have enough income,” said Nicolas Fayon, co-founder and CEO of Jump (pictured above).

Jump wants to support more independent employees in the long run, so it currently offers services to software developers, data engineers, project managers, creative consultants, and sports coaches.

For example, it wants to support B2C merchants, akin to “businesses that bill customers through Stripe, using online payments or physical payment terminals,” Fayon said. Jump also plans to expand to other countries, starting with a U.K. umbrella company for freelancers working within the U.K.

AND latest Financial Times profile Mayayoshi Son opens with the SoftBank CEO seemingly hitting all-time low, gazing his “ugly” face on Zoom and telling himself, “I haven’t done anything to be proud of.”

Indeed, Son largely disappeared from the general public eye after SoftBank’s Vision Fund suffered huge losses on investments like WeWork. But FT journalist Lionel Barber, whose latest biography of Son is titled “Gambling Man,” writes that while Son gave the impression to be “doing penance,” he was in actual fact “plotting a comeback.”

Now SoftBank is specializing in artificial intelligence and has achieved success by taking integrated circuit design company Arm public.

The profile also includes some amusing personal details, resembling Son’s apparent fascination with Napoleon. When the activist investor mentioned Bill Gates and Mark Zuckerberg in a 2020 meeting with Son, he reportedly dismissed them as “one-business men.”

“The right comparison for me is Napoleon, Genghis Khan or Emperor Qin,” Son said. “I’m not a CEO. I’m building an empire.”