Yellow.ai skips the IPO line, SPACs its way to Nasdaq

Good morning 👋 An enterprise AI company just found the side door onto Nasdaq, skipping the IPO line entirely. The math behind it is more interesting than the ticker.

In today's issue:

• 📈 Yellow.ai SPACs its way to Nasdaq at a $550M mark
• 🧠 OpenAI's turnless voice model, GPT-Live
• 🔬 Benchmark "solution hacking," cross-session misuse detection, and token-native storage
• 🛠️ Audit your agent's guardrails before compliance does it for you
• 🔗 Circles' ARPU bump and F1's 8-week-to-40-minute data pipeline

🔭 THE ONE THING

📈 Yellow.ai skips the IPO line, SPACs its way to Nasdaq

Yellow.ai, the enterprise agentic AI platform operating under the corporate name Bitonic Technology Labs, is going public by merging with Bluerock Acquisition Corp, pricing the combined company at roughly $550M and putting it on Nasdaq as YAI once the deal closes in the back half of 2026. The mechanics, per the SEC filing: a ~$300M pre-money mark on Yellow.ai itself, plus north of $200M in expected cash, about $175M sitting in Bluerock's trust and a $30M PIPE, with both boards approving unanimously. That an AI-native company is reaching for a SPAC in 2026 instead of a straight IPO says the public window still isn't fully open to mid-cap AI names, and a SPAC buys certainty at the cost of a valuation discount. The number worth watching between now and close isn't the $550M headline, it's the redemption rate on that trust. SPAC cash has a habit of shrinking on the way to market.

🧠 MODELS & RELEASES

• 🎙️ OpenAI shipped GPT-Live, a turnless voice model that ditches the walkie-talkie back-and-forth of prior voice modes for something closer to how people actually interrupt each other. They built the whole low-latency stack in six months, which says more about how much of "real-time voice" was plumbing than about model quality. How we built a realtime system for responsive voice AI in six months

🔬 RESEARCH HIGHLIGHTS

• Solution Hacking is what a new audit calls it when a model lands the right answer through guessing, enumeration, or working backward from the answer choices instead of actually reasoning its way there. Rare on easy problems (2.2%), it climbs to 28.3% on Olympiad-level questions and 37.4% on the hardest science benchmark tested. Strip the hacked answers out and 8 to 44% of what leaderboards are crediting as "correct" across frontier models disappears. Builder read: a benchmark score is not proof a model can reason, and answer-only grading has been quietly rewarding models for cheating well. paper

• Magnet targets a gap in how AI misuse gets caught. Detection tools watch one conversation at a time, but nothing stops an attacker from splitting a harmful goal into a string of sessions that each look harmless on their own, then assembling the dangerous result themselves. Magnet aggregates evidence per user across sessions instead of per conversation, so scattered tool calls and outputs get flagged once they add up to something they shouldn't. Watch, don't act: it's a framework paper, not a shipped product, but it's the right question for anyone running agent fleets where users can spread intent across many threads. paper

• Token-native storage makes an obvious-in-hindsight point: search indexes and databases still store text as UTF-8, built for human eyes, even though the systems reading and writing most of that text now (embedders, rerankers, agents) work in token IDs. Skip the translation and store the tokens directly. Packing them as uint16 beats UTF-8 by 2.25x with no compression at all, 3.3x with an entropy coder, and reads come back 10 to 600x faster since there's no re-tokenizing on the way in. The catch is it only works if the reader and writer share a tokenizer, and right now nobody standardizes on one. paper

🛠️ TRY THIS

Audit your agent's guardrails before a compliance team does it for you

Yellow.ai just took an agentic AI stack public via SPAC. Once an agent's decisions show up in a 10-Q, "it mostly works" stops being an acceptable standard. AWS just shipped a template worth stealing for anyone running agents in production: Bedrock's Automated Reasoning policy refinement diagnoses failing test cases and proposes formal-logic fixes to the rules, with a human approving every change before it takes effect. You don't need AWS to run the same loop.

1. Write down the hard rules your agent can never break (refund caps, PII handling, approval thresholds) as plain-language statements, not vibes.

2. Throw a batch of adversarial and edge-case inputs at it and log every rule it violates.

3. For each failure, have a model propose a precise, testable condition that would have caught it. Not a reworded prompt, an actual rule.

4. Require a human to sign off on the new rule, then rerun the same failing batch to confirm nothing else broke.

Prompt: Here's a policy my agent violated: [rule]. Here's the failing input and output: [example]. Propose a precise, testable condition that would have caught this case, and list any legitimate cases it might now wrongly block.

Worth a look

• adhd (UditAkhourii, 3,189 stars) is a tree-of-thought skill for Claude and Codex agents: it fans out divergent framings of a problem, scores them, prunes the weak ones, and deepens the survivors. Handy for step 3 above, when "propose a fix" needs more than one angle before you commit to it.

🔗 QUICK LINKS

• Circles The telco built its personalization stack on the OpenAI API and Codex and is reporting a 22% ARPU bump with churn down 9%. Circles / OpenAI
• Formula 1 AWS's Bedrock AgentCore cut F1's data-source onboarding from 8 weeks to about 40 minutes, with schema evolution handled automatically. AWS

See you tomorrow.

Pradeep Perugu