Daily Digest — June 3, 2026
[AI] OpenAI’s o3 Benchmark Claims Are Under Scrutiny
Source: MIT Technology Review · Will Douglas Heaven
The story: Researchers are raising questions about whether o3’s ARC-AGI benchmark performance reflects genuine reasoning or sophisticated pattern-matching on a dataset that may have been partially contaminated during training. The core dispute is whether hitting a benchmark ceiling means the underlying capability is real or just well-optimized for that specific test.
My take: Benchmarks in AI are starting to feel like Goodhart’s Law in real-time — the moment a measure becomes a target, it stops being a good measure. The more interesting question is whether ARC-AGI was even the right test to begin with. I want to understand more about what “reasoning” actually means in this context vs. next-token prediction at scale.
Cloudflare lens: Workers AI serves inference at the edge — if model capability claims are unreliable, the evaluation infrastructure becomes more valuable, not less. AI Gateway’s logging and observability layer is exactly the kind of tooling that lets you run your own evals against real traffic rather than trusting published leaderboards.
[BUSINESS] SaaS Multiples Are Compressing Again
Source: Meritech Capital
The story: Public SaaS revenue multiples have pulled back to 6-8x NTM revenue for median companies, down from the 15-20x peaks of 2021. AI-native companies are trading at a premium but face their own scrutiny on net revenue retention and whether AI features are driving expansion or just inflating headline ARR.
My take: This matters for anyone selling to startups — procurement cycles are tightening and ROI conversations are back. “It’s on the roadmap” no longer closes deals. The question I have: does usage-based pricing help or hurt in this environment? Buyers love the low entry point but finance teams hate unpredictable bills.
Cloudflare lens: Cloudflare’s usage-based pricing model is structurally well-positioned for a budget-conscious buying environment — no seat licenses, no minimums on most products. Worth internalizing as a sales talking point when prospects push back on commitment-based contracts.
[ENG] Postgres 17’s New MERGE Capabilities
Source: Postgres Weekly
The story: Postgres 17 significantly expanded the MERGE statement with RETURNING clause support and WHEN NOT MATCHED BY SOURCE — closing a long-standing gap with SQL Server’s upsert capabilities and making conditional write patterns much cleaner to express in a single statement.
My take: D1 is built on SQLite so the syntax won’t transfer directly, but the pattern of conditional upserts is universal and worth knowing cold. The deeper point is that Postgres keeps eating the “you need a specialized database for that” argument one feature at a time.