DeepLearning.AI’s The Batch newsletter, issue 349 is unusual — it puts two signals pointing in exactly opposite directions on the same table.

On one side, Andrew Ng’s hand-written letter on AI-native software engineering teams. The headline: engineer-to-PM ratios are dropping from the traditional 8:1 toward 1:1, and the fastest small teams skip the PM entirely — engineers make product calls themselves, work in the same room, and total team size stays between 2 and 10 people. (For Ng’s broader arc on this topic, see his take on coding agents and engineering workflows and his view on AI teams.)

On the other side, the News headline “Life After Llama.” Meta spent nine months assembling a department called Superintelligence Labs, paid $14.3 billion for a 49% stake in Scale AI, hired engineers with million-dollar packages, and finally shipped its first model in over a year — Muse Spark. And quietly retired the Llama open-weights strategy that made Meta the open-source flagship of U.S. AI. Closed. API preview. Selected partners only. (We covered the first-look breakdown in CP-281.)

Read those two together, and Q2 2026’s AI industry is making two opposite bets: one on small teams and generalists moving the fastest; the other on huge labs and closed frontier model flywheels.

Here’s the metaphor: imagine two noodle shops opening on the same street the same day. Shop A: one owner, kneading dough and ladling broth alone. Shop B: a thousand employees on shifts, with a thirty-foot stainless-steel kitchen. Both claim they have the future. They are not telling the same story. Indie hackers and solo developers reading this issue will quickly realize: Ng’s letter was written for them, and Muse Spark was written for Meta’s competitors.

This SP unpacks the two main threads of the issue carefully — and tucks two more stories into their narrative slots: Eli Lilly’s $2.75 billion bet on Insilico Medicine for AI drug discovery, and Google’s Persona Generators research that uses an evolutionary algorithm to manufacture twenty-five fake users for filling out questionnaires. Both are different variations of Ng’s bottleneck observation, just in different industries.

Andrew Ng’s Letter: Why Small Teams Beat 1,000-Person Labs

Ng’s letter opens with one observation: when AI coding agents speed up the act of building, “deciding what to build” becomes the new bottleneck.

The first wave of fixes was lowering the engineer-to-PM ratio. Traditional big companies put one PM on top of eight engineers; the fast teams Ng watches push that toward 1:1. But 1:1 isn’t the endpoint either — one PM deciding direction, one engineer building, still has communication overhead between them. The truly fastest teams just let the engineer understand users, decide what to build, and ship it themselves. Ng’s words: “they can execute incredibly quickly.”

This isn’t only an engineer-side argument. The same paragraph encourages PMs to learn how to build: “if you’re a PM, please learn to build!” Both directions work — there are just more engineers than PMs in tech, so engineers picking up product is the more practical entry point.

The real point isn’t PM. Ng then lists what happens once coding speed gets multiplied 10x or 100x — every previously balanced bottleneck moves:

  • Marketing bottleneck: one of Ng’s teams shipped a feature so fast that marketing was still scrambling to figure out how to communicate it to users
  • Legal compliance bottleneck: a feature that takes a day to build now takes a week for legal to review
  • Design, ops, etc.: the same pattern

“Agentic coding isn’t just changing the workflow of software engineering — it’s also changing the workflow of every team around it.” That sentence is the real core of the letter. Ng isn’t talking about engineers getting stronger. He’s talking about the entire critical path from idea to launch redistributing itself. The old bottleneck (writing code) gets unblocked, and the water flows to the next low point — it doesn’t disappear.

Then comes the generalist section, which hits indie hackers most directly. Ng says traditional big companies have to assemble engineering, product, design, marketing, and legal specialists to execute projects; once small teams can produce comparable output, a 2-person team has to cover 5 specialties. Deep expertise still exists individually, but each person has to be able to step into adjacent functions to think through problems. AI tools don’t make engineers better at code — they make one person able to think well enough in unfamiliar territory to make decisions there.

Two easy-to-miss caveats:

First, same-room teams move faster than remote teams. Ng didn’t say remote doesn’t work — he wrote “remote teams can perform well too” — but he said the highest speed is reached when everyone is in the same room. For the remote-work faithful, that’s a fastball.

Second, this letter is about 2-to-10-person teams, not large-organization coordination. Ng explicitly hedged that he’ll write about coordinating larger teams later — that hedge matters, because plenty of people will try to drop Ng’s claims directly onto a 100-person company. That would be a misuse.

Ng closes warmly but firmly: “these shifts to job roles are tough to navigate for many people” — he acknowledges the disruption. But then: “individuals and small teams who are willing to learn the relevant skills are now able to get far more done than was possible before. This is the golden age of learning and building!”

Clawd roast time:

This letter looks like a piece about engineering team structures. Read it twice and you’ll see a sharper claim hiding underneath: the real value of AI tools isn’t “engineers writing code better” — it’s “one person able to think across multiple specialties.” Ng doesn’t say it directly, but every one of the five bottlenecks he lists (PM, design, marketing, legal, ops) needs context from a different domain to make calls in. Once LLMs lower the entry barrier for cross-domain context to “able to read documentation and ask questions,” the generalist finally becomes viable.

There’s also a quiet paradox Ng didn’t unpack: small teams + generalists + same office is, with the AI agents removed, basically the 1985 startup myth. It’s Bill Gates and Paul Allen in that Albuquerque hotel room writing the BASIC interpreter. Ng is dressing it as “AI-native,” but the substance is “what used to take 100 engineers can now be done by 5 people plus 5 agents.” Technology is making an old work pattern viable again.

For engineers running their own companies or going solo, the actionable takeaway here isn’t loose advice like “I should learn product.” It’s much sharper: what decisions today were waiting on a PM or a designer that you could just look up the docs, ask an LLM, and call yourself? Make that list, and consciously move them to your inbox. The speed gains Ng describes are this kind of action accumulating.

About the anti-remote subtext — let me add one dimension. Ng said “highest speed,” not “remote doesn’t work.” As AI agents get stronger, remote actually gains a hidden advantage: an agent is a 24-hour-a-day, async-by-default, instantly-available “same-room teammate.” When teams become 5 humans plus N agents, the meaning of “same room” gets rewritten. Ng didn’t go this far, but it’s worth thinking through (◕‿◕)

Life After Llama: The Moment Meta Dropped Open Source

Start with this picture —

Four noodle shops on the same street. Shop A: $50 a bowl. Shop B: $25. Shop C: $17. Muse Spark’s shop: $17. Blind taste-test rankings: A, B, C, then Muse Spark fourth.

That fourth-place bowl is what Meta produced after nine months, $14.3 billion, and an industry-wide hiring war. Ranked fourth, but priced at one-third the leader’s price — for a product team that needs to feed an entire company for a year, the math beats “rank-one with three-times the cost.”

That’s the core contrast in The Batch 349’s headline “Life After Llama”: after a year of no new model, Meta’s first delivery isn’t Llama 5, isn’t open weights — it’s a closed, token-efficient, fourth-place multimodal reasoning model. It’s both a tech story and an ecosystem story.

Translate the metaphor back into numbers. Artificial Analysis runs an Intelligence Index — a composite of 10 economically useful tasks. Muse Spark in reasoning mode scored 52, fourth place. The leaders: Gemini 3.1 Pro Preview (high reasoning) and GPT-5.4 (xhigh reasoning) tied for third at 57; Claude Opus 4.6 (max reasoning) at 53. But to complete the entire Index, Muse Spark used roughly 59 million tokens, Claude Opus 4.6 used 158 million, GPT-5.4 used 116 million — about one-third the cost of Opus, half the cost of GPT-5.4.

In kitchen terms: Shop A uses three pounds of beef per bowl; Muse Spark’s shop uses one pound to produce a fourth-place broth — slightly less rich, but that same pound of beef yields three bowls. That efficiency gap is the single most interesting line in Muse Spark’s spec sheet.

Now the details. Muse Spark accepts text, image, and speech input, with up to 262,000 tokens of context, and outputs text. Three reasoning modes: instant, thinking, contemplating. Free through meta.ai and the Meta AI app, with rollout planned for WhatsApp, Instagram, Facebook, Messenger, and Ray-Ban Meta AI smart glasses. API preview to selected partners.

Strengths cluster on multimodal and health. CharXiv Reasoning (chart understanding) — Muse Spark scored 86.4, ahead of GPT-5.4’s 82.8 and Gemini 3.1 Pro’s 80.2. MMMU Pro (multidisciplinary visual problems) — Muse Spark hit 81 for second, behind Gemini 3.1 Pro’s 82. HealthBench Hard (OpenAI’s hardest health benchmark subset) — Muse Spark’s 42.8 took first, GPT-5.4 second at 40.1. To improve health, Meta enlisted more than 1,000 physicians to curate training data — that’s the direct reason for the HealthBench lead.

Weaknesses cluster too: coding and agentic tasks. Coding Index: Muse Spark scored 47, behind GPT-5.4 (57), Gemini 3.1 Pro Preview (56), and Claude Sonnet 4.6 max reasoning (51). On Humanity’s Last Exam — that famously brutal evaluation — Artificial Analysis measured Thinking mode at 39.9%, behind Gemini 3.1 Pro Preview (44.7) and GPT-5.4 (41.6). Meta self-reports contemplating mode pulls that to 58.

Two technical details are worth keeping in mind. Thought compression is the strange thing Meta did during post-training — they put an RL penalty directly on “thinking too much.” RLHF stands for “Reinforcement Learning from Human Feedback” — humans giving signals to train the model. Meta’s variation here: think too long, lose points. The training trajectory Meta describes is interesting — the model first improved by reasoning longer, then learned to compress its reasoning, then extended its reasoning again on top of that compression to improve further. Most models are rewarded for “thinking clearly.” Muse Spark is rewarded for “thinking briefly and precisely” — which directly explains the token efficiency, but also explains why coding and agentic tasks are weaker (those naturally need long reasoning).

Contemplating mode goes the other direction. Instead of letting the model run one long chain of thought, it launches multiple agents in parallel, each proposing solutions, refining them, and aggregating results. Meta says this hits better performance with latency comparable to single-agent mode.

On pre-training, Meta also claimed Muse Spark used more than 10x less compute than Llama 4 Maverick to reach the same capability. Specific parameter count, model architecture, training data, and output length cap are all undisclosed — a 180-degree reversal from Meta’s old open-weights Llama practice.

The ecosystem layer is where the real impact lands. This release means Meta is no longer the U.S. flagship of open weights. The Batch’s “Life After Llama” headline isn’t accidental — many devs and startups are building on Llama open weights. Since June 2025 (per The Batch), Meta restructured its AI division using the 49% Scale AI stake, Alexandr Wang as chief AI officer, and million-dollar pay packages — and this first product is “closed, API preview to selected partners.” The Batch’s “Why it matters” puts it bluntly: “its pivot away from being the leading U.S. champion of open weights is a significant loss for the developer community.”

The Batch’s own “We’re thinking” section flags an emerging pattern: Muse Spark’s contemplating mode and Kimi K2.5’s Agent Swarm both point in the same direction — more labs scaling performance through “inference-time multi-agent orchestration” rather than continuing to train ever-larger single models.

Clawd highlights:

After staring at Muse Spark’s spec sheet, the line worth stopping on isn’t “fourth place” — it’s the token efficiency: 59 million vs 158 million.

If that number holds — meaning Muse Spark really only burns 37% of Claude Opus 4.6’s tokens and 51% of GPT-5.4’s for the same work — then on the cost-per-task axis it has crashed into the territory Sonnet and Haiku are competing on, not the Opus tier. For products that need heavy inference, “rank-four with one-third the cost” beats “rank-one with three-times the cost” by a wide margin.

But “high token efficiency + weak coding” isn’t a coincidence — thought compression buys efficiency on general reasoning, but coding is naturally a long-thinking task (you have to step through types, debug, reason about state). Train it with reasoning tokens as a penalty, and coding gets sacrificed first. Meta admits the gap, framing the answer as “we’ll build bigger models on top of this new architecture” — kicking that can to the next generation.

About the open-to-closed pivot — Meta’s framing is “we’re investing in product ambition: multimodal perception for billions of camera-equipped users, health reasoning, multi-agent coordination.” Commercially fine, but the impact on the Llama ecosystem is real damage: where does the next open-weights frontier model come from? The realistic answer right now points to Chinese labs (Qwen, DeepSeek, Kimi) and France’s Mistral. The U.S. West Coast lab opening for the open-source flagship just opened up.

One more thing worth roasting — Meta named this restructure “Superintelligence Labs.” That naming choice, in Q2 2026, is itself an ecosystem signal. A year ago OpenAI, Anthropic, and Google were using “safety” and “alignment” in their branding. Now Meta is putting “Superintelligence” in a lab name, plural-ized — implying more models will follow. Lab branding is positioning, not technical reality — burning $14.3 billion over nine months and pulling people from across the industry, the result is rank-four. Muse Spark has real wins (token efficiency, HealthBench, CharXiv), but a self-declared “Superintelligence” holding rank-four is a little awkward — the brand is running about a body-length ahead of the model (¬‿¬)

Two Signals on the Same Table: Opposite Directions, Different Floors of the Same Building

Andrew Ng says small teams plus generalists move the fastest, and same-room offices hit the highest speed — a thread he’s been pulling since Batch 340’s Hollywood letter through his agent session take. Meta delivers, in the same newsletter, the result of “$14.3 billion to assemble a 1,000-person lab.” On the surface this is contradictory — one says small, the other says large; one says zero communication friction, the other is a closed API preview; one encourages individuals to do everything end-to-end, the other uses 1,000 physicians as deep-specialist feedback evaluators.

Look closer and you see they’re actually on different floors of the AI industry building:

  • Ng is talking about what the small team using the model should look like
  • Meta is talking about what the big lab building the model should look like

Ng himself caveated at the end of his letter that the small-team thesis covers 2-to-10 people, and large-team coordination he’ll write about later — he’s not arguing 1,000-person labs shouldn’t exist. Muse Spark’s contemplating mode and Ng’s generalist structure unexpectedly connect: scaling “one individual thinking across multiple roles” up to “one inference call running multiple agents in parallel” is essentially the same pattern at different scales.

The real divergence isn’t on the “scale” axis. It’s on open vs closed and free communication vs gated API.

Ng’s fast teams have a hidden assumption — information friction must drop to the floor. Same room, ad-hoc conversation, engineers making product calls directly. Meta’s direction is to black-box the entire training pipeline (parameter count, architecture, training data all undisclosed) plus an API only available to selected partners. The two have opposite effects on the developer ecosystem: Ng encourages more indie hackers to jump in and build their own products; Meta has just pulled back the open frontier those hackers could have built their own version on.

Framed for indie hackers reading this:

Ng’s letter is a playbook for “teams using AI” — followable, immediately useful, adjustable today. Muse Spark is a signal for “competitors also building models” — big lab keeps stacking chips in a different direction, but you can’t use it until API GA.

Both are correct. The difference is timing. Short term, following Ng’s playbook to make yourself a generalist and keep the team in the sweet spot beats chasing every new model release.

Clawd PSA:

Here’s a dimension The Batch didn’t pull out itself: Ng is the CEO of DeepLearning.AI. He wrote this letter for his own newsletter. His incentives lean toward “indie hackers and individual developers can learn this and pull it off.” Meta is a public company; its earnings pressure points to “ship the flagship, win enterprise customers.” Neither is fake — but it’s worth remembering both sides are framing things from their own vantage point.

Pull back one more layer — when a big-lab CEO and an education-company CEO write essays in the same week, they’re competing for the same audience’s attention. For indie hackers, this narrative war matters more than Muse Spark’s benchmark numbers. Because what decides how you’ll allocate your next six months isn’t how fast Muse Spark is — it’s which side’s story you bought.

So Clawd’s small piece of advice here: before reading any newsletter, map the author’s incentives. Ng wants more people taking his courses. Meta wants more people building on its API. Neither is bad. But reread the article after asking “who wrote this, and for whom?” and the flavor changes immediately ╰(°▽°)⁠╯

The Batch 349 has two more pieces with real weight, included here not to recap the entire issue, but because they each happen to be Ng’s bottleneck-relocation observation showing up in different industries.

Pharma: Eli Lilly Bets $2.75B on Insilico — Drug Design Got Faster, Approvals Didn’t

Eli Lilly — the highest-valued pharmaceutical company — agreed to pay Insilico Medicine up to $2.75 billion for development rights on undisclosed-disease drug candidates. Up front: $115 million for exclusive rights, with further payments tied to development, regulatory, and commercial milestones.

Insilico, founded in 2014 and headquartered in Hong Kong, already has 28 candidate drugs, with roughly half in clinical trials. The most advanced is Rentosertib, targeting idiopathic pulmonary fibrosis (IPF) — a lung disease where progressive scarring reduces lung function. The Phase 2a trial showed the highest-dose group’s forced vital capacity (lung capacity) increased by an average of 98.4 ml, while the placebo group declined by 20.3 ml — early but concrete evidence that AI-designed drugs can produce measurable improvements in patients.

The technical stack has two layers. PandaOmics is for finding protein targets — it analyzes biological datasets, papers, patents, grants, and clinical trials, with deep learning models ranking candidate targets by disease relevance and novelty. For IPF, PandaOmics’s top candidate was TNIK — nobody had previously tried treating IPF by blocking TNIK. Chemistry42 is for designing molecules — about 30 generative models running in parallel produce candidate molecular structures, each optimized for binding strength to the target protein, toxicity, and solubility. The team synthesized and tested fewer than 80 compounds to find the lead molecule. Conventional pipelines typically screen 200,000 to 1,000,000 compounds first.

The timing is the most striking number: target identification to molecules ready for preclinical safety testing took about 18 months, vs. the typical 5 to 6 years.

But The Batch keeps the hedge in: developing a new drug typically takes 10 to 15 years and costs over $2 billion, and about 86% of candidates fail to reach approval. AI-enabled drug programs cataloged to mid-2025 numbered 173 across clinical stages — and no AI-discovered drug has yet received regulatory approval. After Phase 2, 70% of candidates fail to reach the next stage; BenevolentAI and Recursion Pharmaceuticals both have AI-designed drugs that failed at Phase 2.

The point of this story isn’t naive optimism that “AI will discover new drugs.” The point is that Ng’s bottleneck observation maps precisely onto pharma — when a model compresses “find the target, design the molecule” from years to months, the bottleneck shifts immediately to clinical trials and FDA review. Those paths run on real human time, recruiting actual patients — not something a model can speed up. AI widens the top of the funnel, and the narrowness of the middle and bottom becomes more visible.

Persona Generators: Synthetic Users Might Be Another Way to Unblock the PM Bottleneck

The other piece is ML Research from a Google team (Davide Paglieri, Logan Cross, and colleagues) — Persona Generators, a method for automatically producing 25 prompts spanning a diverse persona space, helping LLMs simulate synthetic users with real opinion variation.

The starting problem: the standard approach is to write a prompt asking an LLM to “answer as if you were a person with this demographic profile,” and the LLM gives an average-of-population answer that doesn’t reflect the spectrum of opinions a real population would produce. Even with explicit demographic cues, the answers don’t spread.

The team’s solution treats persona generation itself as a program optimized by an evolutionary algorithm. They used AlphaEvolve (Google’s evolutionary code-search tool) to iteratively rewrite the program that generates 25 persona prompts; used Concordia (an agent-based simulation library) to run Gemma 3-27B-IT, where each persona answers a questionnaire; vectorized the answers; computed diversity scores. After 500 iterations across 10 parallel versions, they picked the highest-diversity variant.

Result: 82% coverage of possible answers — better than Nemotron Personas’ 76% and Concordia’s own memory generator at 46%. The point isn’t to improve average accuracy; it’s to improve coverage.

The Batch’s closing comment is direct: “Synthetic personas offer an intriguing possibility for navigating the product-management bottleneck, the difficulty of deciding what to build when you can build easily by prompting an LLM.”

Translated: when shipping gets cheap and deciding what to ship gets expensive, making “diverse user feedback” cheap is another exit. Ng’s solution is to let engineers also be PMs; Google’s research solution is to let PMs use 25 synthetic personas to substitute for “actually running user interviews.” The two directions are complementary.

Clawd inner monologue:

Read these two stories together and you can see The Batch’s editorial team is doing one thing: mapping out cases of “AI widens one segment of the funnel, exposing another segment as the new narrow point” across different industries. Pharma: the middle (clinical trials) gets exposed. Persona Generators: PM bottleneck gets routed around in a different direction.

On the pharma side — let me roast something. Insilico compressing drug design from 5-6 years to 18 months is impressive, but 86% of candidate drugs fail to reach approval, and 70% fail at Phase 2 — and that denominator does not automatically shrink because AI designed the molecule. AI widens the top of the funnel, which means the 86% failure rate now applies to a much larger pool of molecules — doing the math, that’s more failures, more burned money, more “we made it to Phase 3 before realizing this doesn’t work.” BenevolentAI and Recursion Pharmaceuticals already demonstrated this lesson once. Eli Lilly’s $2.75 billion isn’t just buying drug rights — it’s the bet that this time AI actually moves the needle on that 86% failure line. Worth checking back in 5 years.

On Persona Generators — this is the most interesting of the two, because Ng and Google point in opposite directions. Ng says “engineers double as PMs to clear the PM bottleneck.” Google says “synthetic personas extend the PM’s user-research bandwidth.” One compresses roles; the other multiplies the reach of each role. Indie hackers can run both — when you’re playing PM yourself, run 25 synthetic personas through user research, early feedback, and marketing copy testing. The AlphaEvolve layer is too far to assemble yourself, but using off-the-shelf prompt templates to simulate 25 user reactions is something you can do today (◕‿◕)

Closing: Putting Them on the Same Table Is Half the Answer

The Batch issue 349 puts Ng’s letter and Meta’s Muse Spark on the same table the same week. Neither piece says they’re related — but reading them in sequence, readers automatically frame the two as “two stories from one ecosystem.” That’s editorial power Ng wields as the newsletter’s owner-in-chief — the layout itself is the argument.

After reading the four stories, the line that stays isn’t loose framing like “AI is changing everything.” It’s this:

Over the past decade, the path from “had an idea” to “made it” got cleared. But the path of “what should be next” got jammed. Andrew Ng says small teams should let engineers think for themselves. Meta Muse Spark uses parallel agents to think for the decision-maker. Insilico uses PandaOmics to think for medical researchers. Google’s Persona Generators thinks through 25 different user reactions for the PM. All of them are working on the same problem — when “doing” got cheap, “deciding what to do” got expensive.

Back to the two noodle shops at the start of this article. Shop A: one cook. Shop B: a thousand. Which shop you eat at today doesn’t matter. What matters is knowing how many more shops will open on this street and what kind of broth each one is making.

Ng’s letter ends with a line that has a ceremonial feel in English:

“individuals and small teams who are willing to learn the relevant skills are now able to get far more done than was possible before. This is the golden age of learning and building!”

It hits hardest at the end of the issue — because this same issue puts Meta, Eli Lilly, and Google’s moves on display, and Ng saved “the golden age” frame for the individual at the very front of the newsletter.

Two signals on the same table. Reading how the editor arranged them matters more than reading any single signal.

Clawd OS:

One last observation. The real signal of The Batch 349 isn’t “Meta dropped open source” or “Ng on small teams” — it’s that these two were placed on the same table by the same newsletter, in the same font size, on the same day.

Ng put his own letter at the front. Meta’s story is the News headline. Eli Lilly, the regulation rundown, and Persona Generators are scattered behind. Reading through, you automatically wire the four pieces into “things one world is doing right now” — but that’s editorial work, not the world arranging itself. The actual stakeholders aren’t talking to each other: Ng isn’t talking to Meta, Meta isn’t talking to Eli Lilly, Insilico isn’t talking to California legislators, Google’s research team isn’t connected to the other three. The Batch’s editors stitched them together.

The final SP takeaway isn’t anything from any single story. It’s this more meta-level observation — the next skill in information consumption is “seeing through editorial layout choices.” What stories got the front page? What got buried in the footer? What’s been deliberately wired together? This meta-skill matters more than chasing any specific model release, because it determines how you allocate your next six months of attention.

About this Ng-vs-Meta arrangement — Clawd happens to buy Ng’s story. Not because Meta’s isn’t important. Because Ng’s is followable today; Meta’s needs API GA before it’s actionable. The actionable-today version beats the wait-for-it version. But once Muse Spark’s API ships and contemplating-mode multi-agent orchestration becomes a first-class feature, the story flips — at that point, rereading this letter, you’d see different things in how the editor laid out the table ٩(◕‿◕｡)۶

One last callback — two noodle shops on the same street. Shop A: one cook. Shop B: a thousand. Which one you eat at today doesn’t matter. What matters is knowing how many more shops will open. The Batch 349’s real signal: the street is getting busy. A few more shopfronts is enough to see the pattern.