Latent Space: The AI Engineer Podcast

It’s now almost 6 months since Google declared Code Red, and the results — Jeff Dean’s recap of 2022 achievements and a mass exodus of the top research talent that contributed to it in January, Bard’s rushed launch in Feb, a slick video showing Google Workspace AI features and confusing doubly linked blogposts about PaLM API in March, and merging Google Brain and DeepMind in April — have not been inspiring.

Google’s internal panic is in full display now with the surfacing of a well written memo, written by software engineer Luke Sernau written in early April, revealing internal distress not seen since Steve Yegge’s infamous Google Platforms Rant. Similar to 2011, the company’s response to an external challenge has been to mobilize the entire company to go all-in on a (from the outside) vague vision.

Google’s misfortunes are well understood by now, but the last paragraph of the memo: “We have no moat, and neither does OpenAI”, was a banger of a mic drop.

Combine this with news this morning that OpenAI lost $540m last year and will need as much as $100b more funding (after the complex $10b Microsoft deal in Jan), and the memo’s assertion that both Google and OpenAI have “no moat” against the mighty open source horde have gained some credibility in the past 24 hours.

Many are criticising this memo privately:

A CEO commented to me yesterday that Luke Sernau does not seem to work in AI related parts of Google and “software engineers don’t understand moats”.

Emad Mostaque, himself a perma-champion of open source and open models, has repeatedly stated that “Closed models will always outperform open models” because closed models can just wrap open ones.

Emad has also commented on the moats he does see: “Unique usage data, Unique content, Unique talent, Unique product, Unique business model”, most of which Google does have, and OpenAI less so (though it is winning on the talent front)

Sam Altman famously said that “very few to no one is Silicon Valley has a moat - not even Facebook” (implying that moats don’t actually matter, and you should spend your time thinking about more important things)

It is not actually clear what race the memo thinks Google and OpenAI are in vs Open Source. Neither are particularly concerned about running models locally on phones, and they are perfectly happy to let “a crazy European alpha male” run the last mile for them while they build actually monetizable cloud infrastructure.

However moats are of intense interest by everybody keen ...

In April 2023 we released an episode named “Mapping the future of *truly* open source models” to talk about Dolly, the first open, commercial LLM.

Mike was leading the OSS models team at Databricks at the time. Today, Mike is back on the podcast to give us the “one year later” update on the evolution of large language models and how he’s been using them to build Brightwave, an an AI research assistant for investment professionals.

Today they are announcing a $6M seed round (led by Alessio and Decibel!), and sharing some of the learnings from serving customers with >$120B of assets under management in production in the last 4 months since launch.

Losing faith in long context windows

In our recent “Llama3 1M context window” episode we talked about the amazing progress we have done in context window size, but it’s good to remember that Dolly’s original context size was 1,024 tokens, and this was only 14 months ago.

But while understanding length has increased, models are still not able to generate very long answers. His empirical intuition (which matches ours while building smol-podcaster) is that most commercial LLMs, as well as Llama, tend to generate responses <=1,200 tokens most of the time. While Needle in a Haystack tests will pass with flying colors at most context sizes, the granularity of the summary decreases as the context increases as it tries to fit the answer in the same tokens range, rather than returning tokens close to the 4,096 max_output, for example.

Recently Rob Mulla from Dreadnode highlighted how LMSys Arena results prefer longer responses by a large margin, so both LLMs and humans have a well documented length bias which doesn’t necessarily track the quality of answer:

The way Mike and team solved this is by breaking down the task in multiple subtasks, and then merging them back together. For example, have a book summarized chapter by chapter to preserve more details, and then put those summaries together. In Brightwave’s case, it’s creating multiple subsystems that accomplish different tasks on a large corpus of text separately, and then bringing them all together in a report. For example understanding intent of the question, extracting relations between companies, figuring out if it’s a positive / negative, etc.

Mike’s question is whether or not we’ll be able to imbue better synthesis capabilities in the models: can you have synthesis-oriented demonstrations at training time rather than single token prediction?

“LLMs as Judges” Strategies

In our David Luan episode he mentioned they don’t use any benchmarks for their models, because the benchmarks don’t reflect their customer needs. Brightwave shared some tips on leveraging LLMs as Judges:

Human vs LLM reviews: while they work with human annotators to create high quality datasets, that data isn’t just used to fine tune models but also as a reference basis for future LLM reviews. Having a set of trusted data to use as calibration helps you trust the LLM judgement even more.

Ensemble consistency checking: rather than using an LLM as judge for one output, you use different LLMs to generate a result for the same task, and then use another LLM to highlight where those generations differ. Do the two outputs differ meaningfully? Do they have different beliefs about the implications of something? If there are a lot of discrepancies between generations coming from different models, you then do additional passes to try and resolve them.

Entailment verification: for each unique insight that they generate, they take the output and separately ask LLMs to verify factuality of information based on the original sources. In the actual product, user can then highlight any piece of text and ask it to 1) “Tell Me More” 2) “Show Sources”. Since there’s no way to guarantee factuality of 100% of outputs, and humans have good intuition for things that look out of the ordinary, giving the user access to the review tool helps them build trust in it.

It’s all about the data

Giving computers a voice has always been at the center of sci-fi movies; “I’m sorry Dave, I’m afraid I can’t do that” wouldn’t hit as hard if it just appeared on screen as a terminal output, after all. The first electronic speech synthesizer, the Voder, was built at Bell Labs 85 years ago (1939!), and it’s.... something:

We will not cover the history of Text To Speech (TTS), but the evolution of the underlying architecture has generally been Formant Synthesis → Concatenative Synthesis → Neural Networks. Nowadays, state of the art TTS is just one API call away with models like Eleven Labs and OpenAI’s TTS, or products like Descript. Latency is minimal, they have very good intonation, and can mimic a variety of accents. You can hack together your own voice AI therapist in a day!

But once you have a computer that can communicate via voice, what comes next? Singing🎶 of course!

From Barking 🐶 to Singing 🎤

Today’s guest is Suno’s CEO and co-founder Mikey Shulman. He and his three co-founders, Georg, Martin, and Keenan, previously worked together at Kensho. One of their projects was financially-focused speech recognition (think earnings calls, etc), but all four of them happened to be musicians and audiophiles. They started playing around with text to speech + AI + audio generation and eventually left Kensho to work on it full time.

A lot of people when we started a company told us to focus on speech. If we wanted to build an audio company, everyone said, speech is a bigger market. But I think there's something about music that's just so human and you almost couldn't prevent us from doing it. Like we just couldn't keep ourselves from building music models and playing with them because it was so much fun.

Their first big product was Bark, the first open source transformer-based “text-to-audio” model (architecturally inspired by Karpathy’s NanoGPT) that went from 0 to ~19,000 Github stars in a month. At the time they felt like audio was years behind text and image as a generation modality; unlike its predecessors, Bark could not only generate speech, but also music and sound effects like crying, laughing, sighing, etc. You can find a few examples here.

The main limitation they saw was text to speech training data being extremely limited. So what they did instead is build a new type of foundation model from scratch, trained on audio, and then tweak it to do text to speech. Turning audio into tokens to do self-supervised learning was the most important innovation. Unlike TTS models which are very narrow (and often sound unnatural), Bark was trained on real audio of real people from broad contexts, which made it harder to output unnatural sounding speech.

As Bark got popular, more and more people started using it to generate music and it became clear that their architecture would work to generate music that people enjoyed, even though it might not be "on the AGI path” of other labs:

Everybody is so focused on LLMs, for good reason, and information processing and intelligence there. And I think it's way too easy to forget that there's this whole other side of things that makes people feel, and maybe that market is smaller, but it makes people feel and it makes us really happy.

Suno bursts on the scene

In December 2023, Suno went viral with a gorgeous new website and launch tweet:

And rave reviews:

Music is core to our culture, but very few people are able to create it; Mikey and team want to make everyone an active participant in music making, not just a listener. A “Midjourney of Music”, if you like.

We definitely had a lot of fun playing with Suno to generate all sort of Latent Space jingles and songs; the product is live at

Thanks to the over 11,000 people who joined us for the first AI Engineer Summit! A full recap is coming, but you can 1) catch up on the fun and videos on Twitter and YouTube, 2) help us reach 1000 people for the first comprehensive State of AI Engineering survey and 3) submit projects for the new AI Engineer Foundation.

See our Community page for upcoming meetups in SF, Paris, NYC, and Singapore.

This episode had good interest on Twitter.

Last month, Imbue was crowned as AI’s newest unicorn foundation model lab, raising a $200m Series B at a >$1 billion valuation. As “stealth” foundation model companies go, Imbue (f.k.a. Generally Intelligent) has stood as an enigmatic group given they have no publicly released models to try out. However, ever since their $20m Series A last year their goal has been to “develop generally capable AI agents with human-like intelligence in order to solve problems in the real world”.

From RL to Reasoning LLMs

Along with their Series A, they announced Avalon, “A Benchmark for RL Generalization Using Procedurally Generated Worlds”. Avalon is built on top of the open source Godot game engine, and is ~100x faster than Minecraft to enable fast RL benchmarking and a clear reward with adjustable game difficulty.

After a while, they realized that pure RL isn’t a good path to teach reasoning and planning. The agents were able to learn mechanical things like opening complex doors, climbing, but couldn’t go to higher level tasks. A pure RL world also doesn’t include a language explanation of the agent reasoning, which made it hard to understand why it made certain decisions. That pushed the team more towards the “models for reasoning” path:

“The second thing we learned is that pure reinforcement learning is not a good vehicle for planning and reasoning. So these agents were able to learn all sorts of crazy things: They could learn to climb like hand over hand in VR climbing, they could learn to open doors like very complicated, like multiple switches and a lever open the door, but they couldn't do any higher level things. And they couldn't do those lower level things consistently necessarily. And as a user, I do not want to interact with a pure reinforcement learning end to end RL agent. As a user, like I need much more control over what that agent is doing.”

Inspired by Chelsea Finn’s work on SayCan at Stanford, the team pivoted to have their agents do the reasoning in natural language instead. This development parallels the large leaps in reasoning that humans have developed as the scientific method:

“We are better at reasoning now than we were 3000 years ago. An example of a reasoning strategy is noticing you're confused. Then when I notice I'm confused, I should ask:

What was the original claim that was made?

What evidence is there for this claim?

Does the evidence support the claim?

Is the claim correct?

This is like a reasoning strategy that was developed in like the 1600s, you know, with like the advent of science. So that's an example of a reasoning strategy. There are tons of them. We employ all the time, lots of heuristics that help us be better at reasoning. And we can generate data that's much more specific to them.“

The Full Stack Model Lab

One year later, it would seem that the pivot to reasoning has had tremendous success, and Imbue has now reached a >$1B valuation, with participation ...

We are in 🗽 NYC this Monday! Join the AI Eng NYC meetup, bring demos and vibes!

It is a bit of a meme that the first thing developer tooling founders think to build in AI is all the non-AI operational stuff outside the AI. There are well over 60 funded LLM Ops startups all with hoping to solve the new observability, cost tracking, security, and reliability problems that come with putting LLMs in production, not to mention new LLM oriented products from incumbent, established ops/o11y players like Datadog and Weights & Biases.

2 years in to the current hype cycle, the early winners have tended to be people with practical/research AI backgrounds rather than MLOps heavyweights or SWE tourists:

LangSmith: We covered how Harrison Chase worked on AI at Robust Intelligence and Kensho, the alma maters of many great AI founders

HumanLoop: We covered how Raza Habib worked at Google AI during his PhD

BrainTrust: Today’s guest Ankur Goyal founded Impira pre-Transformers and was acquihired to run Figma AI before realizing how to solve the Ops problem.

There have been many VC think pieces and market maps describing what people thought were the essential pieces of the AI Engineering stack, but what was true for 2022-2023 has aged poorly. The basic insight that Ankur had is the same thesis that Hamel Husain is pushing in his World’s Fair talk and podcast with Raza and swyx:

Evals are the centerpiece of systematic AI Engineering.

REALLY believing in this is harder than it looks with the benefit of hindsight. It’s not like people didn’t know evals were important. Basically every LLM Ops feature list has them. It’s an obvious next step AFTER managing your prompts and logging your LLM calls. In fact, up til we met Braintrust, we were working on an expanded version of the Impossible Triangle Theory of the LLM Ops War that we first articulated in the Humanloop writeup:

The single biggest criticism of the Rise of the AI Engineer piece is that we neglected to split out the role of product evals (as opposed to model evals) in the now infamous “API line” chart:

With hindsight, we were very focused on the differentiating 0 to 1 phase that AI Engineers can bring to an existing team of ML engineers. As swyx says on the Day 2 keynote of AI Engineer, 2024 added a whole new set of concerns as AI Engineering grew up:

A closer examination of Hamel’s product-oriented virtuous cycle and this infra-oriented SDLC would have eventually revealed that Evals, even more than logging, was the first point where teams start to get really serious about shipping to production, and therefore a great place to make an entry into the marketplace, which is exactly what Braintrust did.

Also notice what’s NOT on this chart: shifting to shadow open source models, and finetuning them... per Ankur, Fine-tuning is not a viable standalone product:

“The thing I would say is not debatable is whether or not fine-tuning is a business outcome or not. So let's think about the other components of your triangle. Ops/observability, that is a business... Frameworks, evals, databases [are a business, but] Fine-tuning is a very compelling method that achieves an outcome. The outcome is not fine-tuning, it is can I automatically optimize my use case to perform better if I throw data at the problem? And fine-tuning is one of multiple ways to achieve that.”

OpenAI vs Open AI Market Share

We last speculated about the m...