7 min read · 12 sources

The API is becoming the legacy interface. An Installmap study of YC startups found that 43 of 119 sampled companies with an MCP server (36.1%) have no public API at all. In the 2025-26 batches, llms.txt adoption has hit 36.2%, surpassing public API adoption at just 22.7%. The next generation of AI-native companies is building for LLMs first, and REST is the thing they’re skipping.

This isn’t a niche experiment. The public API share has stayed flat across batches while MCP and llms.txt have climbed. New startups are treating the LLM as their primary consumer, not a bolt-on. If you’re building a devtools product or an API you expect developers to integrate, the question isn’t whether to support MCP - it’s whether your existing REST endpoint is already the wrong interface.

Perplexity’s Photon engine returns 95% of search results in 230 milliseconds or less, replacing the open-source engine the company used before.

The AI-Native Company Has No Job Titles

Source: elenaverna.com ↗

Field notes from a year inside Lovable describe an organization that has abandoned traditional structure. No job titles, only ICs and Leads. No meetings by default. Roles blur because AI agents handle the routine work, and humans act as “agent parents” supervising automated systems.

The author is honest about the limits: this is one data point, and you can’t separate what’s AI-native from what’s Swedish or unique to Lovable. But the pattern is worth watching. If flat structures with AI agents in the loop become the template, the entire management layer - the people whose job is coordination and status tracking - gets thinner. The takeaway for operators isn’t to copy Lovable, it’s to ask which of your meetings exist because information doesn’t flow, and whether an agent can make that flow automatic.

There Are No More Moats. You Never Had One

Source: auren.substack.com ↗

Auren Hoffman’s argument is blunt: switching costs, proprietary data, and network effects are dead. A friend running a $1B revenue AI startup says he has no moat and must have the best product every week or lose customers. Migrations are free. Contracts are monthly. Customers leave the moment something better ships.

The uncomfortable conclusion is that moats were never the product - they were the migration cost. In the AI era, the only defense is continuous product excellence. For founders, this reframes strategy: don’t build features to lock customers in, build them to win the week. The moat metaphor was always a fantasy; the attack is the only play.

SaaStr's AI Agent: 17,000 Conversations, 600 Meetings, 60% More Business

Source: saastr.com ↗

SaaStr’s inbound AI agent handled 17,000 prospect conversations in 12 months, booking ~600 meetings for its conference. A newer self-serve agent drove a 60% increase in new business. All run by three humans.

The details matter. The agent sits on the highest-intent page - the sponsor page for a ~$90K purchase. It does real qualification: budget, goals, competitors. It books meetings directly in real-time. This replaced a slow, generic form-and-email flow that SaaStr’s own author calls “the worst email on planet Earth.” The lesson for anyone running a B2B funnel: the form is not a funnel stage, it’s a conversion killer. Put an agent where intent is highest, give it a real job, and let it book.

Perplexity's Photon: Search in 160 Milliseconds

Source: threadreaderapp.com ↗

Perplexity’s research account announced Fast Search, running on a Rust-based retrieval and ranking service called Photon. It returns 95% of search results in 230 milliseconds or less, with a median latency of 160ms. Photon now handles all retrieval and ranking across Perplexity, replacing the open-source engine the company used before.

The same thread details a post-trained Computer model with hint-guided self-distillation that cut tool-call failures by 21.2% in a live A/B test, and an open-sourced Unigram tokenizer that reduced CPU utilization by 5-6x. For anyone running search or ranking at scale, Photon’s numbers are the benchmark to beat - and the fact that Perplexity built it with a small team plus hundreds of agents is its own story about how AI-native engineering changes the size of the team you need.

CloudWatch Omni: Observability Without the AWS Console

Source: aws.amazon.com ↗

Amazon CloudWatch Omni is an AI-powered observability experience accessed via a dedicated URL with enterprise SSO - no AWS Console access required. It’s built on OpenTelemetry, so existing telemetry appears without reconfiguration. Omni organizes observability around applications, automatically discovers services, maps dependencies, and adjusts alarms automatically.

The pitch solves real pain: dashboard maintenance, context loss during incidents, and tool-switching. For SREs, the interesting part is the SSO separation - observability becomes a product for the whole org, not just the ops team with console access. And the OpenTelemetry foundation means it’s not a lock-in play in the traditional sense; your telemetry stays portable. The catch is whether the AI alarm adjustment earns trust - auto-tuning alerts is a bold claim for anyone who’s been paged at 3am by a misconfigured threshold.

ElevenLabs Ships Image & Video Generation

Source: threadreaderapp.com ↗

ElevenLabs’ Image & Video API lets you generate images and videos with the same API you use for speech. It supports 18 models, including Veo 3.1, Seedance 2.5, GPT Image 2.5, and Nano Banana Pro. Available today on Pro plans and above, with generations costing the same credits as in the app, and failed generations not charged.

The notable move is the aggregation play - one API key for 18 models across speech, image, and video. For developers, that’s a single integration point instead of juggling providers. The failed-generations-not-charged policy is a small thing that removes a real pain point in working with flaky generative APIs.

Deep Tech Due Diligence Is Getting Worse

Source: postquantum.com ↗

PostQuantum’s analysis argues that venture due diligence for deep tech has declined, particularly for frontier deals where checks are largest. The seed-to-Series A gap grew 84% from 2021 to 2024, and the A-to-B gap grew 97%. The author infers physics checks are being skipped in fast-closing deals, citing quantum as an example with $4.9B in private investment against only $1.4B in market revenue.

This is a warning to LPs and a signal to founders. If technical diligence is being skipped, then deep tech valuations are increasingly based on narrative rather than feasibility. For founders, that’s an opportunity to differentiate by actually doing the verification - a startup that can prove its physics is worth more than one that just claims it.

What It Takes to Build Great Products Now

Source: figma.com ↗

Figma’s take is that AI makes getting to a prototype easier, but the hard work is sustaining speed to launch, choosing the right direction, and creating differentiated products. Strong teams excel in three areas: speed (maintaining momentum through the final 20%), direction (weighing many options before betting), and differentiation (shaping a result only they would make).

The practical advice: use the fastest medium for each change - adjust spacing in Figma, calculate dynamic prices in code. This is the counterpoint to the vibe-coding hype. Prototypes are cheap now; the bottleneck is judgment. The teams that win are the ones that can say “no” to a good direction because a better one exists, and keep shipping through the boring final stretch.

The Messy Middle Is Where AI Money Is

Source: tomtunguz.com ↗

Tom Tunguz’s argument is that the most important AI market is the “messy middle” - multi-step workflows needing a smart-enough model at an affordable price. Anthropic and OpenAI cut prices within 90 minutes of each other, and open models run a majority of token volume at an 86% discount to closed models. Frontier models have fallen from 53% to 45% of corporate token consumption.

Demand is a normal distribution with a fat middle, not a pyramid. The frontier is thin and getting thinner. For anyone building AI products, this means the winning strategy isn’t to chase the smartest model - it’s to find the cheapest model that’s smart enough for the job, and build the workflow around it. The price war at the top is a distraction from where the volume actually is.

Quick Hits

Source: indiehackers.com ↗

• Vacation Tracker hit $3M ARR, and its landing page sat online for 18 months before a waitlist member tracked down the founder to ask when it would launch.
• Good product ideas often die in the pipeline because prioritization, specification, and implementation rely on consensus and intuition - which are poor at discerning good ideas.
• Figma’s advice on direction: weigh many options before betting on one, and shape the result into something only your team would make.