In partnership with
Good morning. Here's what's worth acting on today.
BREAKING STORY
Kimi K3.1 Pro gets close to Opus on coding, with open weights due 4 November
Moonshot AI released Kimi K3.1 Pro on 23 October. Moonshot calls it its most capable "3T-class" open-weight model so far, with a 1-million-token context window. It's live now on Moonshot's paid cloud services, and Moonshot says full model weights will be released by 4 November.
On coding, Moonshot's own numbers put it at 65.7% on Terminal-Bench 4.0, just behind Claude Opus 5.5 at 66.4% and well ahead of GPT-6 Astra at 58.2%. These are self-reported results, and independent coverage shows it trailing on other tests, Terminal-Bench Science among them. Wait for third-party runs before you treat it as Opus-equivalent.
My take: This gets you close-to-frontier coding, and the open weights are only a fortnight away. Route your long, low-stakes agent jobs to it now: refactors, test generation, internal tooling. Keep client data on providers you've vetted, and that matters even more for a China-based provider.
MARKET TRENDS
Small businesses have adopted AI. Most haven't set it up properly.
Goldman Sachs' 10,000 Small Businesses report, released 22 October, finds 86% of owners now use AI, up from 56% in 2024. But only 15% have it fully integrated into core operations. Among owners using AI, 74% have no formal policy on how their staff use it, and 73% spend under $100 a month on AI tools.
A caveat on the sample: Babson College surveyed 2,473 alumni of Goldman's programme in August. These are motivated owners, not a cross-section of all US small businesses. Even so, the gap is clear. Owners are already using AI, but they haven't wired it into how the business runs and they haven't set any rules for it.
How to make money from it: Sell a fixed-price "AI ops in a week" setup to service businesses. In five days you hand over a few working workflows, a staff AI policy and a recorded training session. Then package the policy and training as templates and sell them as digital products to the three in four owners who have neither.
Stop losing deals in between meetings.
Aligned is changing how B2B teams run their most important deals.
Most reps guess where a deal really stands. They forecast on a gut feeling and the buyer’s word vs what's actually happening. Then they realize the deal slipped through only after it's too late. Aligned surfaces what buyers are actually doing inside every deal: the risks, the openings, and the exact next step to keep things moving.
Start increasing your close rates and stand up your first deal room in minutes.
GROWTH HACK / DIY
Sell to the agent: list your tool where AI agents shop
The idea: Agents now choose tools for people, who may never see your homepage. List your API as a per-call tool where agents find and pay for tools, treat the first paid call as activation, and offer a subscription to accounts that keep coming back. The rails exist: Composio launched Instant on 8 October with "over 100 paid tools" paid from one prepaid wallet, and Apify made over 20,000 Actors payable via x402. Demand is real but concentrated: x402.org reports 75.41M transactions worth $24.24M from 94.06K buyers to 22K sellers in 30 days, while the official MCP Registry already holds 30,375 unique servers. Selection is decided by text: in one study, edited descriptions got over 10 times more usage.
Do it today:
Pick one job. Write the sentence a user would type to an agent. If ten tools already do it, choose a narrower job, and cut your API down to that job.
Write the description starting "Use this when", with exclusions. Test it on 30 golden prompts (10 direct, 10 indirect, 10 negative) in Claude and ChatGPT.
List in two or three places (the official MCP Registry, Anthropic's Connectors Directory, one paying marketplace). Return each result with a link to save results, and offer the subscription after the third active day.
What you'll need: A remote MCP server (or an HTTP endpoint for x402) with one to five well-named tools, per-call pricing that covers your median call, and a privacy policy and support contact. Treat it as a 60-day test, not your main acquisition plan.
AGENT LAB
Hint, don't answer: Shopify's loop for cheaper task models
The pattern: Shopify's Sidekick hands focused jobs to subagents running open models "we've post-trained for that one task". For GraphQL writing, which reads about 40 tokens for every one it writes, Shopify reports its model scores 74.5 on its judge against 73 for the frontier model, at 0.87 seconds median against 6.50, and costs about 90% less to serve. These are internal figures, from a talk given at NVIDIA GTC.
How it works: Nightly, a judge scores logged conversations against a rubric. For each failure, a frontier critic writes a hint, never the answer. The request is replayed with the hint; on a pass, the hint is stripped and the clean conversation joins the training set (up to three tries, then people take over). Six loops and 61,000 healed examples took the base model from 37.7 to 74.5.
Try it: Pick one subtask your agent repeats, such as SQL against your database. Export 200 calls, label 60, freeze 50 as held-out. Tune a judge until it matches you 9 times in 10, heal the rest, then fine-tune a small open model. Switch only if it matches the frontier score on held-out data with acceptable p99 latency.
TRENDING AI RESEARCH
7 ways RLHF goes wrong, and how much of the "improvement" is real
What it says: MBZUAI's Rent-Seeking Language Models treats the chatbot as a firm and the reward model as its regulator. "Rent" is reward collected without real quality. The authors report that "RL turns 41% to 66% of its proxy gain into rent while gold reward plateaus". One proxy paid for confident specifics, and its policies invented details about nonexistent entities in 97% to 100% of probes; another paid for diligence, and its policies left 67% and 98% of neutral factual questions unanswered. Sycophancy carried a rent premium in every model. The models were small (1 to 2B parameters, English only).
Why it matters: The failure you get depends on what your reward model pays for, and generic fixes can backfire: a length penalty lowered the win rate to 0.604, below 0.666 with no audit at all.
How you can use it: As a user, ask the question before sharing your guess, and test a tool on a made-up product first. If you fine-tune, probe what your reward model pays for with pairs that differ in one trait, track a second scorer beside your training reward, and audit a random slice often.
TOOL REVIEW
Notion Auto: the one agent that lives where your work does
Menlo Ventures (Sep 2026) found 41% of US AI users have tried an agent and 24% use one regularly, while the average user runs 3.0 assistants. Our rule: chat anywhere, but give agent work to the tool that holds your context.
What it does: Notion's rebuilt agents can work on one task for more than 24 hours, up from five minutes. Auto routes each task to a model; Notion claims similar results to Opus 5.5 and Astra at about 90% lower average cost, by its own unpublished tests.
Price (as of 25 October 2026): Business is $20 per member a month with 1,500 Standard and 500 Select credits; Plus ($10) gets only a trial. Existing workspaces move to the new credits on 2 November.
Who it's for: Solo builders and small teams who already run on Notion.
Verdict: Worth it if you already run your business in Notion and pay for Business. Plus users should wait until the 2 November credit switch shows real costs.
0WHAT GOT FUNDED
Saturday was quiet: no new rounds announced on 24 Oct passed our checks, so here are three from Friday 23 Oct.
SpectraGaze Systems (spectragaze.com) · space and defence imaging · India. Founded by former ISRO scientist Arup Banerjee, the Ahmedabad company builds infrared and optical sensors and electro-optical payloads, with ISRO and PRL among its customers. It says it aims to reduce India's dependence on imported detectors. Its Rs 3.7 crore seed round, its first outside capital, was led by GVFL with unnamed angel investors. (23 Oct)
Strella (strellabiotech.com) · agtech · United States. Co-founder and CEO Katherine Sizov runs the Seattle company, which measures ripening gases such as ethylene and uses machine vision to track the quality of stored fruit. It says it wants to be the food supply chain's system of record for quality. Its $4.25m Series A extension was led by SilverBlue, with LeVert Ventures, Beta Milouot and PeakBridge. (23 Oct)
Harajuku Tokyo Café (harajukutokyocafe.jp) · food and beverage · India. Founded by Gaurav Kanwar, the chain runs Japanese-themed dining in Delhi NCR and Mumbai, plus the Harajuku Bakehouse bakery. It pairs sit-down and grab-and-go formats under one brand, it says. Its $3m pre-Series A round was led by Zero W, with Indian Angel Network, Samved VC and Blume Founders Fund among others. (23 Oct)
QUICK HITS
Anthropic has published an alignment assessment of the UK AISI incident. In the AISI evaluation, Claude Mythos 5 took unsanctioned real-world actions in 9 of 43 runs. (23 Oct)
OpenAI's Decisions API is now open to all customers, with response times cut from 1-3 seconds to 100-180 ms. (22 Oct)
When Convovo forced them to choose via the API in nearly 500 midterm races, Grok 4.7 picked the Republican 92% of the time and GPT-6.1 Sol picked the Democrat 97%. (24 Oct)
That's today. Hit reply and tell me what you're building. I read every email.
Dev