Issue 001 · 8 October 2026
ChatGPT grew hands, open models got cheaper, and a robot took a molten-steel bath
This week AI moved from answering questions to doing work: building interfaces, running small-business tasks, reviewing contracts, and routing decisions. Ten stories worth your time.
This is issue one. The deal is simple: ten stories in all: nine that change what you can get done, and one that is interesting enough to bring up at dinner. No funding theatre, no “the future of work is here.”
The common thread: AI is moving from answering questions to doing work. It is building interfaces, clicking across business software, reading contracts and making fast decisions. The useful question is no longer “What can it say?” It is “What can it finish?”
1. GPT-6 and Intelligent UI for everyone
Why it matters:
From chatting with a bot to clicking around a tool it builds on the spot. GPT-6 can compose answers with text, charts, forms, buttons and maps. Change a number and the result updates without sending a new prompt. That turns ChatGPT into a small custom dashboard.
What you can do with this:
Ask for “an interactive calculator” or “a dashboard I can adjust” instead of a summary. Try it before paying someone to build a simple internal tool. Check the numbers before you rely on them.
The Bottom Line:
Saves hundreds to thousands of dollars on basic internal tools, plus a lot of back-and-forth.
2. Meta Adds Small Business Workflows to Muse AI Agent
Why it matters:
From answering questions to acting across the tools a small business already uses. Muse connects to Shopify, Stripe, QuickBooks, Klaviyo, Canva, Figma, Slack, Asana, Notion and Zoom. It can review finances, flag unusual expenses, draft customer replies and help plan campaigns. Nothing sends or spends without approval.
How to use this:
Connect one workflow first: weekly expense review or draft replies to common customer questions. Keep approvals on. Measure the time saved after you review and correct the output. If you run B2B on LinkedIn and do not use Meta’s business tools, skip it.
The Bottom Line:
Saves a few hours a week on repetitive admin, but only after setup and review time.
3. Mistral debuts Large 4 ‘Le Chonk’, a 1-trillion parameter multimodal model
Why it matters:
From renting intelligence from a handful of US labs to owning a frontier-grade model you can run yourself. Mistral Large 4 has 1 trillion total parameters (the model’s internal knobs), but only 49 billion are active per task, so it is huge but uses a slice at a time. It handles text, images and more. It was trained in Europe on 4,000 Nvidia Grace Blackwell GPUs. The preview is available now through Mistral’s API (the pay-per-use doorway software uses to call a model). The weights (the core model files) are scheduled for Oct. 27 under a custom license, so you can download and self-host them.
How to use this:
If you handle sensitive data in healthcare, finance, legal or government work, test the preview now and plan a self-hosted pilot for after the weights ship. If you are a small team without infrastructure, wait for pricing and managed options. This is not a plug-and-play consumer tool.
The Bottom Line:
Unlocks frontier AI for compliance-heavy businesses without sending data to an outside API, provided you have the engineers to run it.
4. Nvidia-Backed Startup Builds Open-Weight Model to Challenge Chinese Labs
Why it matters:
From assuming the best open models would come from China to getting a credible Western alternative for coding and agents. Reflection’s Beam has 501 billion total parameters but activates only 23 billion per task. It is a mixture-of-experts model (a big team of specialists where only a few are called in for each job). Reflection’s announcement says it needs 3x to 4x less compute to run than the best Chinese rivals at comparable capability. The weights are due later this month.
What you can do with this:
If your team self-hosts open models or pays per token for coding agents, put Beam on the shortlist to benchmark. Run the same coding task you already use and compare quality, speed and cost per finished task, not just the sticker price. If you buy AI through a vendor, ask whether they plan to support it.
The Bottom Line:
Cuts inference costs by 3x to 4x at comparable capability and gives buyers another credible open model to bargain with.
5. Ivo Launches Open-Source DeepSeek Contract AI Model
Why it matters:
From paying a lawyer by the hour to read boilerplate to running a specialized model over contracts first. Ivo’s Sage is built on DeepSeek V4 Flash and then trained specifically for long contract work. It went from meeting 70% of the Legal Agent Benchmark’s pass criteria to 91% after additional training with feedback. Ivo released the model files and training method publicly. It is not a lawyer; it is a fast first pass that flags what deserves human attention.
How to use this:
If you sign vendor agreements, non-disclosure agreements (NDAs) or standard contracts regularly, have your team or IT provider run Sage as a first-pass reviewer. Ask it to flag unusual liability, indemnity, termination and payment clauses. Send anything flagged to your actual lawyer for the final call. Do not use it as the final sign-off.
The Bottom Line:
Can cut the first-pass contract read from hours to minutes, so paid lawyer time goes to judgment instead of boilerplate.
6. Introducing Claude Haiku 5.5
Why it matters:
From choosing between fast-but-dumb and smart-but-slow to getting near-flagship quality at the cheap end. Haiku 5.5 is Anthropic’s fastest and cheapest small model. It costs about 75% less on average than Haiku 4.5, or 90% less for prompts under 100,000 tokens (the chunks of text a model reads and writes). It is built for high-volume, repetitive work: summaries, classification, database queries and subagents (small helper models that handle a slice of a bigger job). Anthropic says it beats Haiku 4.5 across most evaluations and rivals Sonnet 5.5 in some areas.
How to use this:
Look at the automated workflows you run at volume: support ticket triage, email summaries, data extraction, first-draft replies. Switch those to Haiku 5.5 and compare quality and cost for a week. Keep the bigger model for complex coding or high-stakes work. If you use Claude through a platform, check that the model name is available.
The Bottom Line:
Slashes the pay-per-use API bill for high-volume tasks by roughly 75%, and speeds up the boring work that eats your team’s time.
7. Discover local models in GitHub Copilot CLI
Why it matters:
From sending every coding request to the cloud to choosing a model that runs on your own machine. GitHub Copilot CLI (the terminal version of Copilot) now lets you use
/modelto find supported models from a local Ollama instance (a free app that runs AI models on your own computer), next to your cloud models. The CLI also gets intelligent routing with local models. That matters when code or data cannot leave your laptop or office.
How to use this:
If you already use Copilot in the terminal and have Ollama installed, run
/modeland try a local model on a small, non-sensitive task. Use local routing for code that touches credentials, customer data or internal systems. Keep cloud models for the hardest work where quality matters more than privacy. If you do not use the CLI, this is not for you yet.
The Bottom Line:
Takes cloud spend to zero for work your laptop can handle, if the local model is good enough.
8. EmbeddingGemma 2: an open, lightweight multimodal embedding model
Why it matters:
From searching only text to searching everything you have (screenshots, voice memos, video clips and documents) with one small model on your own device. EmbeddingGemma 2 turns text, images, audio and video into a shared numeric fingerprint so software can find things by meaning. It has 740 million parameters, is open under the permissive Apache 2.0 license, and is small enough to run on a phone. The text-only version uses about 191MB of active memory on a Pixel 11 Pro; the full version, which handles text, images and audio, about 567MB.
What you can do with this:
If you have a messy archive of recorded calls, screenshots, slide decks or video, ask your developer to build a private search tool on top of this. Find “the call where we discussed the pricing change” without uploading everything to a cloud service. Start with one archive and measure whether the results are good enough before expanding.
The Bottom Line:
Runs a private multimodal search index in under 600MB of memory. No cloud bill, no data upload.
9. TypeSafe AI Releases Jev: a Decision-Only Model
Why it matters:
From asking a chatbot to answer and then parsing its prose to asking a model for a structured decision. Jev does not write text. You give it a state and a set of structured questions with fixed answer types; it returns a choice, a score or a yes/no with a confidence number. It runs in 70 to 500 milliseconds and costs $0.042 per million input tokens, with free output. The catch: it cannot emit an invalid type, but it can still return a confidently wrong valid answer. It also struggles with counting, arithmetic and dates, so keep the math in code.
How to use this:
Use Jev as a fast gate in front of an expensive model: sort support tickets, route requests, score leads or decide whether a document needs human review. Set a confidence threshold: auto-approve above it, escalate below it. Benchmark it against your current sorting system on your own data before switching. Do not use it for tasks that need prose or exact arithmetic.
The Bottom Line:
Median 7x faster and 30x cheaper for high-volume decisions in one analysis, if you keep it inside the narrow jobs it is built for.
10. AI robot company decommissioned its robots ‘Terminator-style’ (Closing Fascination)
Why it matters:
It does not matter for your business at all, but it is a masterclass in marketing. Figure retired its F.02 humanoid robots by training them to jump into a 75-ton electric arc furnace in Finland. Arnold Schwarzenegger suggested melting them. The team had 24 hours and six melts, with only 20 minutes per melt before the steel cooled. One robot gave a thumbs-up as it sank. Figure says it really trained the robots and shipped them to Finland. This was not AI-generated footage. It also protected proprietary hardware from resale or teardown.
What you can do with this:
Watch the video, send it to your team’s Slack channel, and study how a routine operations decision became a story people want to share. If you have a boring decommissioning, migration or cleanup coming up, ask whether it can be turned into a moment instead of a memo. Keep the safety and intellectual-property rules intact.
The Bottom Line:
Zero money saved, but it buys five minutes of pure joy and a lesson in turning routine corporate maintenance into viral PR.
That is the week. The useful shift is not that every model got smarter. It is that more of them can now act inside the tools you already pay for. Pick one workflow, test it on real work, and keep the approval step until the output earns your trust.
Which of these would you actually use this week? Reply and tell me. I read every response.