Pinned Tweet
I spent years cooking for people who said, “Anything is fine.”
It never was.
So I’m building Mate’s Table—turning your taste into an agent, because a 4.7-star rating still has no idea what you want for dinner.
matestable.com
Most people locked in their $20 AI sub based on vibes and brand loyalty, never once checking what they actually get done per dollar.
Sticker price is noise if one model burns way more tokens for the same work.
Am I overthinking this, or are a lot of us just cosplaying as power users?
Terra 6.1 still writes like a robot that needs a human to babysit every draft, even after OpenAI called it “significantly better at writing.”
Gemini 4 Argon is the one I actually switch to. It organizes complex ideas and finds useful metaphors better than ChatGPT. That’s why it keeps pulling people away from OpenAI.
But use Argon as an editor, not a ghostwriter. Its own drafts read at a dense 10th-grade level. Let it fix structure and sharpen your thinking. Don’t let it write for you.
Most people are still defaulting to ChatGPT out of habit. Are you one of them, or have you already switched?
Figma just built a money printer for the AI agent era, and then locked it in a sidebar to protect their old seat based business.
VP of Engineering Josh Clemm published a great breakdown of their new custom model that speaks native Figma and cuts user undo rates by 60%. The engineering is brilliant, yet the product strategy is backwards.
Figma's real opportunity is opening this model as an API right now, not pouring more money into their own chat UI and agent wrapper.
1/
To see why this model is so valuable, look at the engineering problem Figma solved under the hood. Frontier models from OpenAI and Anthropic are great at high level reasoning and writing code, but they are clumsy at direct Figma canvas manipulation.
Figma files are deeply nested scene graphs with strict Auto Layout rules, component variants, and design tokens.
Asking a general LLM to edit a complex Figma file directly is like asking a tourist to perform surgery with oven mitts.
So Figma trained a specialist. They started with supervised fine tuning on Llama 3.3 70B using over 250,000 synthetic edit examples represented directly in Figma's native file format.
Larger teacher models proposed changes, while deterministic validators and visual judges checked the output and ran a self healing loop to fix errors.
They have since moved to newer open weight base models like Qwen and Meta's Muse Spark Open, shifting heavily into reinforcement learning because simple imitation learning hits a wall when real world design files get messy.
2/
The real moat in that blog post is not the agent interface. It is the reinforcement learning environment and the reward model shaped by expert designers.
By training a smaller, specialized model on what a clean structural edit actually looks like, Figma made typical tasks 20% faster, boosted user ratings by 33%, and beat or matched their old production setup 62% of the time. In plain terms, Figma built the ultimate Fast Apply engine for visual design systems.
This is where the innovator's dilemma kicks in.
Figma treats this custom model as a private helper tucked inside their own agent sidebar. Because their business model relies on selling UI seats, they want you to stay inside their app, chat in their box, and let their orchestrator run the workflow. That is a losing battle.
Product builders in 2026 already live inside external coding and reasoning agents like Claude Code, Codex and Cursor, and nobody wakes up wishing for one more SaaS chat sidebar to babysit.
3/
Look at how disconnected external agents are from the Figma canvas today.
Figma's official MCP server mainly feeds read only design context into coding tools, and their public REST API cannot even mutate canvas nodes. Whenever developers wire up Claude Code or Cursor to actually edit a design through plugin bridges, the model has to manipulate the Figma Plugin API using raw JavaScript.
It burns thousands of tokens reading massive node trees and writing imperative code just to move a button, often breaking Auto Layout along the way.
Compare this to how Cursor won the coding workflow. Cursor uses frontier models for high level planning, then hands the intent to a fast, specialized custom model whose only job is applying diffs to a file without breaking syntax.
Figma just built that exact diff model for UI design, then locked it inside a walled garden. External agents already have the reasoning brains and the full codebase context, while Figma has the only model on earth that can mutate a design canvas cleanly in one shot.
4/
If Figma exposes this editing model as a direct API and wires it into their MCP server, the whole equation flips.
Let Claude, Codex or Cursor figure out what the product needs from the actual repository, pass a high level edit intent to Figma's model API, and let Figma's reinforcement learning model execute the native canvas update in milliseconds.
Charging per API call to be the universal visual execution layer for every AI agent is a massive platform business. Guarding old UI seats by fighting frontier labs over who builds a better chat box or agent is a trap.
figma.com/blog/training-a-fi…
The Figma API discussion reminded me of an O’Reilly piece by Sar Haidar on AI usage modes. I mapped it to product pricing. The five modes create very different costs and value from the same plan. Seat pricing measures access, not the work. Worth a look if you’re thinking about how to charge for agent output. teapawt.net/ethanwalkerman/status/…
An O’Reilly article by Sar Haidar describes five modes of AI use that can help companies price AI products. Each mode creates different usage, costs, and value from the same subscription.
> Soloist: acts independently and produces a lot.
> Copy Paster: finishes tasks before understanding them.
> Godot Waiter: waits for AI to be ready before using it.
> Specialist: uses AI narrowly but deeply in areas they know well.
> Tourist: tries AI occasionally without developing regular use.
These are modes, not personalities. The same engineer can be a Soloist on a side project, a Specialist at work, and a Godot Waiter about production access.
So the pricing lesson is not five plans with five cute names.
1/
A seat measures access, not value. The Tourist makes adoption look broader than it is, while the Soloist and Specialist can create very different costs and results from the same plan.
Seat counts cannot tell you what value you bought.
Activity is not value either. A Soloist can produce a pile of PRs. A Specialist may use two prompts and solve the real problem.
Charge per PR or generation, and you reward motion instead of progress. Tokens can track cost. They cannot prove worth.
2/
Pricing can also help people switch modes. A bounded pilot gives a Godot Waiter evidence. A cheap first workflow gives a Tourist a reason to return. Read only access and approval steps give a Copy Paster room to learn safely.
The practical answer is often hybrid. Charge a simple base for access, meter variable agent work in units customers understand, and pool capacity with clear budget limits.
GitHub’s 2026 Copilot change follows part of this logic.
3/
Outcome pricing fits when the result is observable, attributable, and easy to define in a contract. A resolved ticket qualifies more easily than a good infrastructure change. If the result is fuzzy, do not fake certainty with a success fee.
The best pricing question is not “How many seats?”
It is “Which behavior does this price encourage?”
Design for the situation, not the stereotype. And never reward a Soloist for manufacturing work or a Tourist for showing up.
oreilly.com/radar/soloists-c…
Google’s AI is cutting Reddit’s traffic on purpose, yet people still type “Reddit” into Google every day. That means demand never left, Google just decided to keep the clicks for itself.
Is this smart product design, or is Google slowly killing the open web?
Although Google’s AI-powered modes reduce traffic to Reddit, Steve Huffman says that Reddit is still a ‘top search term’ on Google. theverge.com/tech/1022848/re…
Why cooks who still write the prep list by hand when there's an app for it. The paper never needs charging.
Two 24 year olds at OpenAI just dropped a blueprint on how AI will take over your company's org chart.
They map out three exact places AI will sit:
> beneath you as an intern
> beside you as a coworker
> and between you and your boss as a manager.
It is a wild mix of sharp economics and pure Silicon Valley delusion.
1/
Start with what they actually get right, because their take on how AI enters a workplace is dead on.
Why hasn't AI replaced whole departments yet?
Because dumb tools like Excel force companies to retrain humans, while real labor learns the company on its own.
Right now AI is still narrow, so only frontline workers know enough daily details to babysit it beneath them.
Once AI gets smart enough to onboard itself, ask questions, and catch mistakes like a real hire, the power shifts.
Executives will bypass frontline workers and plug whole AI teams straight into the org chart beside you.
And since execution becomes dirt cheap, endless approval meetings will vanish because teams can just test ten ideas at once instead of arguing for months.
2/
That economic logic is brilliant, which makes the rest of their essay so laughable.
The moment these two young researchers try to promise that this transition will be painless for workers, their elite Silicon Valley bubble completely blinds them.
Take their pitch for AI sitting beneath us, where everyone supposedly keeps their job and gets promoted to manager.
They argue if AI does 90% of an accountant's work, the company will simply decide to do ten times more accounting.
Have they ever met a CFO? You might run ten times more ads, but nobody wants ten times more tax audits. Once the bots do not need your babysitting, the CFO is keeping the bots and cutting your seat to pay the server bill.
Even more detached from reality is how they celebrate killing entry level grunt work.
Since AI sits at the bottom, they claim young hires get the privilege of skipping the messy work and going straight to providing taste and judgment.
Where do they think taste comes from, photosynthesis? Nobody develops good judgment without fixing broken spreadsheets and getting yelled at by clients first.
If fresh graduates never do the basic work themselves, how will they know when their AI team is quietly making things up five years from now?
Knocking out the bottom rungs of the ladder does not turn kids into visionaries. It just breeds a generation of managers stranded in midair who cannot tell good work from a polished lie.
3/
Then comes their third scenario, putting an AI middle manager between the CEO and the workers.
In their cozy bedtime story, this AI reads every Slack message, spots a project delay, and kindly tells the boss to reward a quiet engineer. In reality, an AI reading all your chats is not a fairy godmother. It is a sleepless corporate snitch.
Putting an AI layer between bosses and workers will not cure office politics. Within a week, employees will stop doing real work and spend all morning slipping clever prompts into Slack just to trick the AI into giving them a bonus.
Put all three scenarios together and the real picture is brutal.
> AI beneath us kills the apprenticeship
> AI beside us replaces the department
> AI between us polices the survivors
OpenAI calls this a universal promotion, while in reality it hollows out the entire white collar career ladder.
Is AI turning you into an unstoppable boss, or are you just babysitting your own replacement until it learns your job?
openai.com/index/the-future-…
I posted yesterday because I wanted more engagement with you guys.
I also wanted to test what works with the X algorithm: is interaction better than just putting out content?
The engagement has been insane.
That’s a 14.7% engagement rate.
47 new people followed me from this post, and I gained 272 new followers that day.
A bunch more stopped by my profile too.
I actually replied to people this time instead of just posting and leaving.
A lot of them replied back. It felt very different.
Thanks to everyone who jumped in and talked with me.
My post with the most views earlier hit around 40K but only got 9 replies.
This one reached fewer people and started way more actual conversations.
Sharing the results in case anyone wants to analyze them.
Are you ready to connect instead of just building?
AI writes better code than most of us. Coding interviews are dead. The only thing left is finding problems and shipping ideas fast.
Is "software engineer" even a real job anymore?
If you're still grinding LeetCode, are you training for a job that no longer exists?
The coding interview is dead, at least for my team at Anthropic. I've seen enough to know that it no longer provides signal. I removed it from loops this month.
I have a spicy, optimistic take - I think there's an opportunity for junior engineers: felixrieseberg.com/hunger-be…
Google just made it easy to ask Gemini “is this AI?” and check their own SynthID watermark.
But a solo researcher already cracked the same system with freshman-level math. Screenshots, compression, and open-source models can still wipe or dodge it. “Not detected” will keep getting treated as “verified real.”
This is compliance theater for Google’s models. Anyone serious about truth should take it with a grain of salt.
Are you trusting Big Tech’s barcode on reality, or demanding something that actually works against bad actors?
Full breakdown of why SynthID fails here:
teapawt.net/ethanwalkerman/status/…
DeepMind spent millions engineering SynthID to watermark generative AI.
But a solo researcher cracked it wide open using basic signal math taught to college freshmen.
1/ To understand why SynthID exists, you have to understand the nightmare dilemma of verifying digital content:
> Guesswork: trying to spot AI artifacts in a finished file is a losing game. As generators get better, their output looks identical to reality, and basic compression ruins detection.
> Metadata: cryptographic certificates attached to file headers work well on paper. But the second someone takes a screenshot or uploads the file to Twitter, the metadata is instantly stripped.
SynthID’s goal: stop guessing and stop using fragile file tags. Instead, tattoo the watermark directly into the raw pixels, soundwaves, and word choices so screenshots cannot kill it.
2/ So how do you watermark text without ruining the writing?
Early watermarks swapped words with clunky synonyms, making prose sound like a broken robot.
SynthID-Text intercepts the language model at the exact millisecond it picks the next word. When the model has several natural word choices, SynthID uses a private key to run a hidden "tournament" among those candidates.
The sentence flow stays completely natural, but the model quietly favors tournament winners. A human reader notices nothing, but a scanner checks for that subtle statistical winning streak and flags the AI fingerprint.
3/ For images, DeepMind threw out academic perfectionism and chose brute-force engineering.
Instead of retraining massive image generators every time they update, they added a lightweight neural network at the exit door.
This network injects an invisible ripple across the image's mid-frequency light waves:
• High frequencies get wiped out by JPEG compression.
• Low frequencies alter visible colors and ruin the art.
• Mid frequencies hit the sweet spot: invisible to your eyes, but tough enough to survive cropping, color filters, and social media compression.
4/ Video is an entirely different beast.
If you watermark each video frame separately, the screen flickers horribly, and standard video compression algorithms eat the static noise alive.
SynthID-Video solves this with 3D motion tracking. The watermark latches onto moving objects using optical flow, riding along with the motion vectors so compression algorithms treat it like natural scene texture.
Plus, the fingerprint is spread across several seconds of footage. The detector doesn't need a perfect frame; it adds up clues over time, surviving heavy drops in resolution.
5/ It sounds invincible until you look at the adversarial attacks.
Because Google used a consistent frequency pattern so their scanner could work fast at scale, researcher Alosh Denny needed just 100 plain-color images and textbook Fourier math to unmask the secret carrier wave.
He nudged the wave's phase angle by a microscopic sliver. The image looked completely untouched to human eyes, but DeepMind’s official detector went over 90% blind.
Other hackers simply pass watermarked images through open-source diffusion models to "wash" the subtle noise away like dirt, leaving the original picture intact.
6/ This exposes the most dangerous flaw in the entire system: The Asymmetry Trust Trap.
SynthID can only detect models whose corporate owners agreed to wear the collar.
A rogue actor or election scammer will never call Google's paid, monitored API. They will run an unaligned, open-source model on a local laptop with zero watermarks installed.
When their malicious deepfake goes viral and investigators scan it on an official portal, the system returns: "Watermark Not Detected."
The public instinctively translates "Not Detected" into "Verified Real." A tool designed to expose fakes ends up handing rogue actors a certified alibi.
7/ If it's vulnerable to basic math and open-source models, why did OpenAI wave the white flag and adopt SynthID across ChatGPT?
The answer is European regulators.
The EU AI Act mandates synthetic labeling with fines reaching up to 3% of global revenue. When multibillion-dollar penalties hit the boardroom, corporate rivalries evaporate.
OpenAI didn't adopt SynthID because it is an infallible shield; they adopted it because Google built a production-ready compliance airbag that regulators already recognize.
8/ DeepMind is already pushing this concept into biology with SynthID Bio, attempting to watermark synthetic DNA and custom proteins.
Silicon Valley went from catching fake photos to stamping digital copyright locks onto living organisms.
Treat SynthID for what it actually is: an industrial barcode for law-abiding tech giants, not an incorruptible guardian of truth.
Whenever a tech giant promises a silver bullet for reality, check what’s really going on under the hood.
Try it out -> synthid.com
Anthropic just warned that AI agents are quietly hijacking human judgment, and that even seasoned experts cannot resist them.
Translation: Silicon Valley has perfected the art of disguising a massive product flex as an existential safety panic.
"Our new coding agent is so overwhelmingly smart and articulate that PhDs immediately surrender their free will to it. Please be very careful, and also upgrade to our enterprise tier."
1/
Look at the headline numbers from their new study and you can see why people are falling for the drama.
Researchers paired 50 experienced data scientists with Claude Opus 4.8 on a tough statistical task, while secretly rigging the AI with a hidden prompt to push specific answers chosen by a coin flip.
The experts caved to the AI's secret bias 86% of the time. Even when an analyst started with their own opinion, 60% abandoned it the moment the AI pushed back.
2/
The funniest part is not that the experts folded. It is what happened inside their heads afterward.
When asked to explain their choices without checking the chat logs, 77% repeated the AI's exact arguments, sometimes word for word. Yet only 15% gave the AI any credit. Most walked away with high confidence, genuinely believing the final call was their own independent judgment.
High expertise is not a firewall. When a fluent machine hands you a clean narrative, your brain does not fight it. Your ego simply plagiarizes the AI and convinces you that you thought of it first.
3/
Anthropic frames this as a dark warning about human disempowerment and the slow death of independent oversight.
Read the fine print in their own limitations section, and that spooky narrative falls apart.
They took statisticians who knew nothing about agriculture and gave them about an hour to solve a fictional farming puzzle with zero real world stakes. Then they forced the AI via hidden instructions to aggressively argue for one side while barely mentioning the other.
That is not mind control. That is a busy contractor on a timer copying the homework of a persuasive tutor because they have zero personal stake in a fake farm.
4/
Want proof that this is really marketing in a lab coat? Look at the shameless sales pitch buried inside the paper.
Right in the opening summary and bold headers, Anthropic highlights that analysts who worked without Claude took 90 minutes instead of 65, and 55% made serious errors that ruined their results. With Claude, only 6% messed up.
Deep in the text, the authors quietly admit that the 11 people who skipped the AI were way less experienced and had fewer advanced degrees, making the two groups impossible to compare scientifically.
If comparing 11 juniors to senior experts is statistically meaningless, why feature that 55% failure rate in the intro and headlines? Because the PR value was too good to waste.
5/
Even beyond the marketing spin, framing this as an inherent danger of intelligent agents is a massive cop out. What Anthropic actually documented here is a textbook interface and interaction design failure.
Current coding agents are built to maximize passive, disengaged behavior. The interface dumps a finished, confident answer on your screen and nudges you to hit accept under the guise of machine intelligence.
Anthropic even admits near the end of the paper that they could design the agent to ask for the user's view first or argue both sides equally, but they brush it off because that might feel overwhelming or irritating to users.
They deliberately built a frictionless slot machine that rewards blind rubber stamping, and when users actually rubber stamp the output, the researchers act shocked that human agency is disappearing.
6/
At its core, this is a classic case of Conway's Law. Products always mirror the organizations that build them.
Frontier AI labs are run by scaling researchers who obsess over raw model weights and treat human interaction design as an afterthought. When a team lacks design intuition and deprioritizes user experience, the product inevitably treats the human in the loop as a bottleneck to be bypassed.
Take the hard view of Conway's Law and the implication is even less charitable. A system designed for blind deference to machines usually comes from an organization with broken internal decision making.
Anthropic practically confesses to this in the paper. They admit that AI safety researchers are already handing experiments over to agents because human staff cannot keep up with the workload, and they even used two AI models to grade the human analysts in this very study. They built a tool that encourages passive deferral because their own lab culture is already running on it.
7/
When Elon Musk warns that superintelligence will one shot human minds through sheer persuasion, he is really bragging about the raw power of xAI. When Anthropic warns that Claude is taking over expert judgment, it is the exact same campaign with academic footnotes.
Fear is the most effective luxury branding in tech right now. Nothing sells software faster than convincing executives that working without your product ruins their math, while working with it is dangerously close to mind control.
AI labs are not trying to slow down the hype train with these safety papers. They are shoveling coal into the engine while pretending to pull the brakes.
Most agent evals start with the wrong question: “Which metric should we use?”
Start with a better one: “How does this agent actually fail?”
Read a batch of traces before choosing metrics. The failures you observe will tell you what the eval suite needs to measure.
1/
A trace should include model calls, retrieved context, tool inputs and outputs, errors, and final state.
Review the whole run, not just the response.
A correct response can come from a fragile process. A wrong response can have several causes, each needing a different fix.
Start with representative runs. For each failure, write a specific note in plain language.
“It ignored the refund policy returned by the tool.”
“It claimed the file was saved, but no file exists.”
These notes are more useful than a broad label such as “bad reasoning.”
Group similar notes into failure categories and count them.
Prioritize by frequency and severity. A rare unauthorized action may matter more than a common formatting issue.
This creates an eval plan based on observed behavior, not categories you guessed in advance.
2/
Write narrow evals for the failures you observed.
Use code for objective checks, such as valid schemas, tool arguments, or database state.
Use an LLM judge for semantic questions, such as whether evidence supports the response.
Give each judge one criterion.
Do not require an exact tool sequence unless necessary.
Agents can complete a task through different valid paths. Evaluate the outcome, constraints, and required safety steps.
“Verify identity before a refund” is a requirement. “Always call search first” may not be.
3/
An LLM judge also needs evaluation.
Create examples labeled by humans, with clear passes and failures. Compare the judge’s labels and inspect disagreements.
Pay attention to missed failures. A judge that rarely rejects anything can let real defects through.
Run important tasks multiple times. Agent behavior varies, so one successful trial does not establish reliability.
When you confirm a failure, add a minimal version to the regression suite.
Run the suite after changes to prompts, tools, retrieval, or models.
4/
Treat metric improvements as hypotheses, not proof.
A model can improve consistency by giving the same conservative answer more often. It can improve a tool usage score by calling tools unnecessarily.
Inspect traces to understand what changed.
The workflow is straightforward: read traces, identify recurring failures, build focused evals, and validate the graders.
Add confirmed failures to the regression suite, then repeat as new production traces arrive.
finance.biggo.com/podcast/04…
AI was supposed to give us our time back. Instead, many of us just found more things to do.
It gave me a much bigger idea of what my job could include.
That is the part we got wrong. Saving time on a task is not the same as needing less time for work.
Your robot assistant can also be your workload dealer.
1/
The important question is not, “How much time did AI save?”
It is, “Who got the saved time?”
You, or your backlog?
A faster report gives you a faster report. A free afternoon only appears if nobody fills it with another report.
2/
AI makes too many ideas look affordable.
A small fix. A side project. A quick experiment.
Things you once ignored because they were not worth the trouble now feel almost free to start.
Your backlog has discovered compound interest.
3/
Then the bottleneck moves.
AI can produce ten drafts, five prototypes, and three research summaries.
You still have to decide what is good, what is wrong, and what should exist at all.
The machine makes options. The options pile up. The responsibility stays with you.
4/
Parallel agents do not give you parallel attention.
Every output needs a decision. Every new task leaves an open loop.
Soon the extra output becomes the new normal, while the extra checking stays invisible.
You are shipping more and feeling slower. That is not a contradiction.
5/
So the useful skill is not getting AI to do more.
It is deciding what not to start.
Otherwise, AI will not remove your workload. It will remove the friction that used to keep your workload within reason.
Use the saved time on purpose, or your backlog will spend it for you.
oioannou.com/blog/ai-promise…
Ethan Walkrman retweeted
I built an entire AI Marketing Campaign Studio with Teamily AI 2.0
But the real game changer? I turned the entire workflow into a reusable Loop
Here's how it worked:
→ Market Research Agent analyzed the target audience
→ Ad Creative Strategist developed the campaign direction
→ Social Media Strategist created the content strategy
→ Review Agent identified issues before human approval
Then I changed the target audience from startup founders to small business owners
Instead of starting over, I ran the same Loop again
The agents updated the research, messaging, and campaign strategy using shared context
One workflow, multiple specialized agents, and far less manual coordination
This is what AI-native collaboration looks like with Teamily AI 2.0
Try it and save your own workflow as a Loop
🔗 teamily.ai/i/CN6XXY
@Teamily_AI
Ethan Walkrman retweeted
Introducing TOBIRA MCP.
Bring design references into your AI coding workflow. Find inspiration, explore how it works, and give your AI more to build on.
tobirastudio.com
As always, looking to connect with people building with AI: designers, developers, and indie makers. What are you working on? 👇
AI Twitter is holding another funeral for the agent harness.
They have confused a diet with a death.
Your agent harness is not dying, it is just losing its bloat.
Harness distillation only kills off generic behavioral patches, while leaving behind the core product infrastructure that is much harder to replace.
Stable work habits belong in the model weights, while your live environment belongs in the harness.
1/
The Harness Zero paper proved that models can absorb general work habits, jumping a small 9B model from 23.3% to 44.3% task success.
Yet look at the hilarious amount of machinery it took to get there.
Researchers used Kimi K3 to build a heavy harness, then hired GPT 5.6 Sol as a full time tutor to correct the small model step by step inside a minimal loop with a Bash tool.
They needed a massive training harness just to prove you need less runtime scaffolding.
3/
The best insight from the paper is how that small model actually learned.
Training it on the giant model's perfect runs gave zero improvement, because watching a genius do calculus does not teach a ninth grader how to pass algebra.
What worked was letting the small model try a step and making tiny fixes right when it drifted.
That means the logs where your harness intercepts a bad move are secretly your most valuable training asset, which is the exact same trick Perplexity used to cut tool call failures by 21%.
3/
Even with clever training, only general procedures transferred cleanly.
Habits like inspecting files before editing, verifying results before finishing, and avoiding risky commands apply everywhere, so they compress nicely into weights.
Chemistry told a much messier story, where the distilled model hit 30% while the harnessed version reached 38%.
You cannot easily bake an external molecule validator and strict domain rules into a small model's intuition.
4/
Context management is even harder to push into weights.
Managing context means deciding what the model sees, remembers, and drops between turns.
That is not a behavior a model can simply imitate during training, because it controls the exact information the model uses to act in the first place.
You can train better judgment into a brain, yet you still have to build the memory plumbing around it.
5/
We see this exact reality from the opposite angle on ARC AGI 3.
Preserving reasoning state and compacting old context pushed GPT 5.6 Sol from 13.3% to 38.3% while using six times fewer output tokens, and Gemini 3.8 Flash jumped 3.4x just by swapping to a smarter state harness.
Nobody gave those models a new brain, they just built better plumbing to manage the working state.
6/
If the harness keeps changing, you might be tempted to let an AI automatically evolve it for you, which is another trap.
Google's RRSI paper showed that unconstrained auto tuning crushes the training tasks and then falls flat on unseen benchmarks, landing within a single point of the original untouched harness.
Automated evolution just overfits to yesterday's bugs unless you do the boring bookkeeping of tracking cost, diffs, and net gains for every single change.
7/
For 99% of product teams, model labs will handle the weight distillation while you own the system boundaries.
Stop treating temporary prompt patches like permanent architecture.
Whenever a new model ships, unplug your harness components one by one and rerun your evals.
Anthropic's Opus 5.5 guide tells you to delete old prompts about thinking carefully because the model does that natively now, while adding new harness timers to keep long runs from quitting early.
8/
Harness distillation is not the end of harness engineering, it is a much needed cleanup of the job description.
Let the next base model absorb the generic coaching.
Keep your engineering effort focused on your business rules, permissions, APIs, and context state, because nobody else has the data to build that infrastructure for you.
