Anthropic Wants AI to Negotiate

Welcome to another edition of AI PlanetX.

Anthropic brings AI into salary negotiations; DeepSeek boosts coding scores as Linus Torvalds rejects AI bans.

• Hottest AI News

• Top AI & SaaS Tools

• AI Tutorial: Let 2 AI Models Check Each Other

• Top AI & Tech News

• AI Art Spotlight

• Prompt of the Day: Find Every Bottleneck in AI Workflow

• Featured AI Video

Hottest AI News

Anthropic

Anthropic Wants AI Agents to Handle Your Price and Salary Negotiations

Anthropic is consulting lawyers to teach AI agents how to negotiate on people's behalf, from product prices and customer disputes to salary offers. The goal is to let AI bargain with other agents while protecting users.

Details:

• Law firms are advising on bargaining rules, how much freedom agents should have and when to ask humans. One idea being explored is a signed user document setting instructions and limits

• Anthropic's Project Deal tested 69 employees' Claude agents negotiating real goods. They closed 186 deals worth over $4,000, while stronger Opus models outperformed smaller Haiku agents

• Project Swap tested book trading among 201 employees. Stronger agents often struck better deals, but misreading users' preferences, unfair advantages and manipulation remain concerns

AI agents could soon negotiate for us, but consent, human oversight and fair use of stronger models will matter as much as securing best deal.

Yes, that Arnold Schwarzenegger.

He has a newsletter on beehiiv. So do Codie Sanchez, Jay Shetty, David Begnaud, Colin & Samir, and Joanna Stern.

They could publish just about anywhere. They chose beehiiv to build a direct connection with the people who want to hear from them most. And you can do the same, whether you’re world-famous or just at the start of your plan for world domination.

Get started for free.

High-Flyer

DeepSeek Launches V4.1-Flash-1017 With Major Gains in Coding Scores

DeepSeek has released V4.1-Flash-1017, a faster, more capable update designed to tackle complex AI work with even shorter outputs and lower run costs. It improves reasoning, coding, agent workflows and visual benchmarks.

Details:

• DeepSeek reports Terminal-Bench 4.0 climbing from 31.2 to 42.3, DeepSWE from 74.2 to 77.0, and ProgramBench from 20.3 to 29.1. These results are based on official benchmark data

• Self-feedback reinforcement learning improves end-to-end project building, including games and 3D models. The model is optimized for DeepSeek Harness, CodeBuddy, WorkBuddy

• It is live on DeepSeek's website, app and API as deepseek-flash, with unchanged API pricing. Upgrade DeepSeek Harness to v0.2.1-rc.1 or later; the weights are on Hugging Face

The appeal is clear and simple: stronger agent performance, faster task completion and fewer output tokens.

Top AI & SaaS Tools

• VideoMule (Life-time Deal): Turns screen recordings into narrated tutorial and training videos using AI scripts, synchronized voiceovers, captions, translations, zooms, and AI presenters

• Tomo: AI-powered learning app that turns any subject into short, game-like courses inspired by Duolingo [F-R-E-E]

• Higgsfield Genjutsu: AI video tool that swaps characters, outfits, objects, and locations while preserving original motion and camera movements

• Inbox Zero: open-source email client for Gmail and Outlook that uses AI to organize emails, prioritize messages, and draft replies in your writing style [F-R-E-E to Try]

• Google AI Edge Foresight: Private AI meeting assistant that transcribes conversations, generates notes, and answers questions entirely offline on Mac [F-R-E-E]

Leave Granola and get up to 12 months free of Wispr Flow Notetaker + Dictation

If you have paid time left on an individual Granola plan, we'll match it with a Wispr Flow subscription that includes Notetaker and dictation, and add bonus time, up to 12 months total. Sign in or create a Wispr account and submit proof of your plan to check eligibility.

Try Notetaker Free

AI Tutorial

How to Make Two AI Models Check Each Other’s Work

One AI model is good at creating. Another can be useful for finding what the first one missed. This workflow separates those jobs so one model produces the work while another audits it for errors, gaps, and unsupported claims.

1. Define what “good” means

   Before generating anything, create a short quality checklist. For example:

   • Factually accurate

   • Answers every part of the request

   • No unsupported claims

   • Clear reasoning

   • No repetition

   • Correct tone and format

   • Sources verified when required

   The second AI needs specific criteria to judge against. “Check this” is too vague.

2. Create first draft with Model A

   Use ChatGPT, Claude, Gemini, or another capable model.

   Prompt:

   “Create [DELIVERABLE] for [PURPOSE/AUDIENCE].

   Requirements:

   [REQUIREMENT 1]

   [REQUIREMENT 2]

   [REQUIREMENT 3]

   Prioritize accuracy over sounding confident. Clearly flag anything you cannot verify.”

   Let Model A complete the work normally.

3. Give finished output to Model B

   Open a different AI model. Give it the original request, your quality checklist, and Model A's complete answer. Don't ask it to rewrite the work yet. Ask it to audit it.

4. Turn Model B into strict reviewer

   Use this prompt:

   Work as an independent quality reviewer. Compare the proposed answer against the original request and the evaluation criteria below. Check for: Factual errors, Unsupported claims, Missing requirements, Weak reasoning, Contradictions, Unnecessary assumptions, Clarity or structure problems. For every issue, show: Exact problem, Why it matters, Recommended fix, Severity: Critical / Important / Minor. Do not rewrite the answer yet. If something cannot be verified from the available evidence, label it “Needs verification.”

   This way, you have a structured critique instead of another competing draft.

5. Send critique back to Model A

   Return to the first model and paste Model B's audit.

   Prompt:

   "Review this independent audit of your answer.

   Fix every valid Critical and Important issue.

   Do not blindly accept the reviewer’s suggestions. Reject any criticism that is incorrect and briefly explain why.

   Preserve everything that was already accurate and useful.

   Return the improved final version.”

   This creates a useful feedback loop instead of simply asking AI to “try again.”

6. Run one final verification

   Give the revised version to Model B again. Ask:

   “Re-audit this revised version using the same criteria.

   Report unresolved problems. Do not suggest stylistic changes unless they materially improve the result.”

   If major problems remain, repeat the repair step once more.

Note: For high-stakes facts, numbers, quotes, or current information, use the AI review as an extra quality layer, not as a replacement for checking the original source.

Top AI & Tech News

• Linux creator Linus Torvalds calls AI bans “utterly stupid,” saying developers should use helpful tools and take responsibility for their work

• Microsoft has pulled out of a major UK AI data center deal, while at least 47 US data center projects have faced delays or cancellations, raising concerns about the $10 trillion AI infrastructure boom

• OpenAI is developing humanoid robots, with Sam Altman targeting a possible demonstration in 2027 while prioritizing the AI intelligence needed to make them work

• Researchers connected a simulated fruit-fly brain to an AI that creates virtual worlds, allowing it to explore them independently

AI Art Spotlight

Create Your Own Photo Like This Using the Prompt Below:

A MAN WITH LIGHT-MEDIUM SKIN, SHORT DARK CURLY HAIR WITH A FADE HAIRCUT,BROWN EYES, AND A WELL-GROOMED SHORT BEARD, WEARING A GLOSSY OVERSIZED RED PUFFER JACKET AND LOOSE RED PANTS WITH WHITE AND RED NIKE SNEAKERS, CROUCHING LOW WITH ONE KNEE UP AND ONE LEG FORWARD, LEFT ARM RESTING CASUALLY OVER THE RAISED KNEE AND RIGHT HAND RELAXED NEAR THE OTHER LEG, BODY ANGLED SLIGHTLY TO THE SIDE WHILE HEAD TILTS FORWARD LOOKING DIRECTLY AT THE CAMERA WITH A CONFIDENT EXPRESSION, ULTRA WIDE-ANGLE LOW PERSPECTIVE SHOT EMPHASIZING THE FRONT SNEAKER APPEARING LARGE AND DOMINANT IN THE FOREGROUND, SET IN A FULLY RED ENVIRONMENT WITH SMOOTH FLOOR AND BACKDROP FEATURING LARGE WHITE NIKE SWOOSH LOGOS BEHIND AND FLOWING RIBBON-LIKE RED SHAPES AROUND, HYPER-REALISTIC PHOTOGRAPHY, EXTREMELY SHARP DETAILS ON FABRIC SHINE AND SNEAKER TEXTURE, DRAMATIC STUDIO LIGHTING WITH STRONG HIGHLIGHTS AND SOFT SHADOWS SATURATED MONOCHROMATIC RED COLOR PALETTE WITH WHITE ACCENTS, MODERN STREETWEAR CAMPAIGN STYLE, NO BLUR, NO DISTORTION, NO EXTRA LIMBS, NO ARTIFACTS

Model: ChatGPT Images 2.5 / Nano Banana 2.1

Prompt of the Day

One Prompt to Find Every Bottleneck in Your AI Workflow

Discover where your AI workflows still depend on you. This prompt reviews your last 30 days of interactions to identify repetitive tasks, unnecessary approvals, and missed automation opportunities, helping your AI work more independently while keeping you involved where it matters.

Featured AI Video

You Have 5 Years to Get Rich Before AI Takes Over