Struck down in June. Back at $103,265.

Unify Your Teams and Tech Stack With HubSpot

Connect your customer data, teams, and tools without the hassle of complex integrations or lengthy setups. One easy platform gives marketing, sales, and service teams a unified customer view and the tools to turn it into growth.

Why HubSpot and what's new
• Generate leads and automate marketing with Marketing Hub
• Build your pipeline and close more deals with Sales Hub
• Scale customer support and drive retention with Service Hub
• Keep customer data clean, connected, and actionable with one, unified platform

Join 306,000+ in over 135 countries using HubSpot to grow their businesses.

See what a more connected approach to growth can do for you and your team. Get setup quickly and start checking off your hardest tasks.

Get started free

Welcome back technologists🫡

The $100,000 H-1B fee that a federal judge ruled unconstitutional in June is back, re-priced at $103,265 per petition. The rest of the stack is the eval discipline behind agents that keep 6,000 users' trust. This week I wired a second model into our own pipeline because the first kept grading its own homework, so the theme hits close to home (about 4 minutes).

Here's the signal today: 🛸
👨‍💻🐕️ (ayo denk)

🧾 DHS re-prices the H-1B petition at $103,265.
🧪 The eval desk keeping 6,000 sales-agent users.
🧪 Your memory format can swing scores 72 points.
🧲 Runable raises $21M to fix the after-the-build part.
🧲 A 4-bit model that beats its full-precision parent.
🐒 Video AI breaks; monkey brains don't flinch.
✈️ Why the next Moonshot may not happen in the U.S.
📌 Libel threats as a response to facts.

🧾 The fee is back

The fee is back. DHS published a proposed rule on August 25 charging $103,265 for every cap-subject H-1B petition, payable at filing, on top of all other applicable fees. A federal judge ruled the earlier $100,000 version unconstitutional in June for violating the separation of powers, and an appeals court affirmed. The trap is reading that ruling as the end of the story: the court struck down a presidential proclamation, so the administration re-filed the same number as agency rulemaking, per JURIST, and the comment window closes September 24. Universities, certain nonprofits like hospitals, government roles, and renewals are exempt; everyone else pays, against a 65,000-visa annual cap plus a 20,000 advanced-degree exemption. If a hire you are planning involves a petition, reprice the offer before the rule finalizes. 💰 Money play. Visa cost is now a line item the size of a seed check.

🧪 Agent Watch

The eval desk behind 6,000 users. Sait Izmit of Snowflake shipped go-to-market agents to 6,000 users, and the build story he co-authored shows the machinery: more than 5,000 user queries a week, automated tests on approved prompts, and real-world checks that combine sampling, manual review, and LLM judges. Snowflake's own Agent GPA judges caught 95 percent (267 of 281) of human-annotated errors on their test set, against 55 percent for the baseline. A sales agent that invents a number loses the account quietly; the eval desk exists so it never gets the chance.

Your memory format is a score variable. The RENDER benchmark fixed the conversation and only changed how the history gets rendered: summary, typed record, or raw log. Across 500 LongMemEval questions and nine models, matched-budget resolved packets beat recency-truncated raw dialogue by 42.4 to 72.6 points, and ChatGPT-style entries scored higher than raw conversation on seven of nine models, with per-model significance mixed. Before you ship memory, hold out 50 questions and A/B the template. 💰 Money play. Token bills and churn both pay for sloppy inputs. An eval bank and a render test are the cheapest insurance on your next deploy.

🧲 The Build

Sell the after, not the build. Runable raised $21M in an all-equity Series A co-led by Susquehanna Venture Capital and Nexus Venture Partners, at a $65M post-money valuation, with a 15-person team in Bengaluru and a $2M annualized revenue run rate three weeks after launching payments. Its agent builds websites, apps, and presentations; the new money extends it into the grow side, where Runable aims to run ad campaigns, manage social media, handle SEO, and optimize how a business shows up in AI chatbot results. If you sell to small businesses, the sellable SKU is growth, not another builder.

Shrink first, heal second. Multiverse Computing's recipe: compress the architecture, quantize to 4 bits, then run a healing step to recover the damage. Their latest result beats its own full-precision original on seven of nine benchmarks, and open-weight releases like gpt-oss, NVIDIA's Nemotron family, and their own Hypernova 60B already rely on compress-then-heal. For a solo deployment, smaller hardware bills without the usual accuracy tax. 💰 Money play. Runable says the margin moved to distribution; Multiverse says it moved to post-training. Your moat is ops, not model access.

🐒 Frontier

Motion outlives makeup. Comparing humans, macaque visual cortex, and neural networks, researchers found most video models generalize poorly when an object's appearance changes, even when motion structure is preserved. Predictive world models held up best and matched primate IT cortex most closely, though no model reproduced the brain's shift from appearance-first to motion coding. If your pipeline touches video, bias toward world-model architectures and test under appearance shifts before you trust it; brittle video models burn GPU hours on babysitting.

✈️ The talent exit

The founder visa that isn't. Rest of World's Lex Zhao argues America is pricing out its future AI champions. The administration is weighing a $100,000 fee on Optional Practical Training, the one-to-three-year work bridge roughly 419,000 people were working on in 2024, and the post-degree grace window has been cut from 60 days to 30. His case study: Yang Zhilin, Carnegie Mellon PhD, internships at Google and Meta, an Apple offer, who went home and built Moonshot AI, whose Kimi model now draws comparisons to OpenAI and Anthropic. 💰 Money play. Incorporation and hiring jurisdiction are now line items. Price them like the rest of your stack.

📌 The Docket

Trump's personal lawyer threatened the Center for American Progress with a $5 billion defamation suit unless it retracted a report finding his National Guard deployments failed to cut urban crime, per NOTUS. CAP refused: "Truth is not and cannot be defamation." If you publish research, dashboards, or criticism with numbers attached, legal noise is now a standard cost of distribution. Document everything.

🎓 Level Up: the agent trust gate

Run these five checks before any agent ships:

Freeze an eval bank pulled from real user asks. Never let it touch training.
Run automated tests on approved prompts every deploy, the way Snowflake's go-to-market desk does.
Score real-world runs with sampling, manual review, and LLM judges; Agent GPA-style judging catches 95 percent of annotated errors while a baseline judge catches 55.
A/B your memory render (summary vs raw log vs typed record) on 50 held-out questions.
Log every failure with the exact packet that caused it, and feed it back into the bank weekly.

Hit reply and tell me which check your agent fails first. If this issue saves you a bad deploy, forward it to one builder who ships agents: nomadsignal.ai

See you tomorrow,
Chase & Kobe 👨‍💻🐕

One tap. Feedback goes straight to the desk.