Studios in the room
5
plus founders, technical directors and engineers
01
AI FIRST · LONDON · 6 AUGUST 2026
The Debates.
The demos were good. The arguments were better. Eight things a room of game and creative studios could not agree on.
Tencent Broadgate Tower · with Supercell, Sumo Digital, Opus Artz, Lucid Games and friends
THE ROOM
Builders, not spectators.
Live demos
6
shipped work, not slideware
Debates left open
8
the reason this deck exists
Everyone presenting had something running in production. Everyone watching had a reason to doubt it. That is what made the arguments worth writing down.
02 / 13
THE HEADLINE ARGUMENT
Data Informs. Taste Decides.
Research said no to a mobile franchise entry. Roughly 300 million new-to-franchise players said yes.*
The taste camp
Games ride on magic that historic data cannot predict.
Chasing the data is trend-chasing: by ship date the trend has already moved. The winning move is a no-compete moat built three to six years ahead of where the data points.
The risk named in the room: slick automated tools can make managers accept AI output without critical thinking. Mediocrity by default.
The data camp
Taste without evidence is just an expensive hunch.
Analysis catches the blind spots a studio cannot see from inside its own pitch, and makes the reasoning behind a bet legible to everyone who has to fund it.
Nobody argued for less data. The fight was about who holds the pen at the moment of decision.
Where it landed
Data should inform the big creative bets, never decide them. AI is a sounding board, not the greenlighter.
*Case discussed in the room, shared under first-name attribution only. Figure is new-to-franchise players, not total franchise players.
03 / 13
HOMOGENISATION
Same Tools, Fifty Identical Games.
If every studio pre-games its pitching and green-lighting with the same tools, what comes out the other end?
The homogeneity risk
Shared tools converge on shared answers.
Feed the same market data through the same models and you get fifty games that look, code and feel identical. The tool optimises toward the safe average, and the safe average is where everyone already is.
The sharper version: AI does not make bad games, it makes the same game.
The internal-value counter
The same tool is genuinely useful behind the door.
For multi-year project alignment it earns its place: guarding against scope creep, catching team desync, keeping a hundred people pointed at the same target across years.
The distinction is not the tool. It is whether it runs before submission or before invention.
Where it landed
Use AI as a stop-gap and an objective sounding board before submission. Its home is alignment and verification, not original ideas.
04 / 13
TRUST AND VERIFICATION
Most Community Data Is Noise.
Everyone wants to mine player sentiment. Almost nobody trusts what comes back.
The problem
Volume is not signal.
Gaming-community data is mostly noise, and a model asked to summarise noise will produce a confident, fluent, wrong answer. The output looks the same whether it is right or not.
The failure mode is not the hallucination. It is that a hallucination reads exactly like a finding.
The fixes on the table
Filter the input, then judge the output.
Connect only to pre-filtered, verified sources, then run an LLM-as-a-judge layer that checks every citation strictly matches the claim it supports. Weight sources by credibility: a heavily-liked review outranks a thousand drive-by posts.
The room also riffed on black-hole analysis: reconstruct the thing you cannot observe from independent signals around it.
Where it landed
Verification is not a feature you add later. It is the thing that decides whether the output is usable at all.
05 / 13
WHERE THE MONEY GOES
Five Principal Engineers, Or Unproven AI Tooling?
Half a million pounds, one budget line, and no agreed way to measure the return.
Hire the humans
Predictable output beats an unmeasurable promise.
Five principal engineers have a known cost and a known throughput. AI tooling has an uncertain cost curve and a productivity gain nobody in the room could put a number on.
The honest question underneath: if you cannot measure it, how do you defend it at the next budget review?
Buy the leverage
Headcount scales linearly. Tooling does not.
The counter was never spend it all on tools. It was roll out slowly: creative leads and directors first, building training plans and learning the workflows before any mass spend.
Sequence beats size. A small deployment that teaches the org is worth more than a large one that lands on people who will not use it.
Where it landed
Nobody won this one. The measured middle was phased adoption through the people who set the standards, not a big-bang rollout.
06 / 13
THE ECONOMICS
Every Re-Prompt Is A Tax.
The honest question for any team: right now, which one is bigger?
The AI tax
What you pay to fix wrong output.
Every correction, every re-prompt, every round of coaxing the model back onto the task. It is invisible on the invoice and very visible in the calendar, which is exactly why it goes unbudgeted.
It compounds quietly: the tax is highest on the people least equipped to spot a wrong answer.
The AI dividend
What you keep when it lands first time.
The hours saved when the output is right, and the work that simply would not have been attempted at the old cost. Frontier models cut the tax and speed the dividend, which is the case for paying more per token, not less.
Cheaper models can be the more expensive choice once the correction time is counted.
The practical control
Put cost metrics on the dashboard. Visible spend keeps teams inside budget and nudges more efficient behaviour without a policy.
07 / 13
THE CULTURAL WALL
Coders Adopted. Artists Did Not.
Same company, same tools, same training. Two completely different answers.
Why coders said yes
Their core work was never the typing.
A programmer's value sits in architecture and judgement, so handing over repetitive code costs them nothing they were proud of. The craft stays where it always was.
Adoption was fast because nothing identity-shaped was being asked for.
Why artists said no
The craft is the job, and the brand is the risk.
Artists protect the craft fiercely, and studios fear what happens to a brand if players learn AI touched the original design. One team went as far as asking for a scanner to detect and strip AI-generated assets before ship.
Read it as a business risk assessment, not stubbornness.
The opening
Artists welcomed AI for the un-precious middle: gray-box placeholders and blockout biomes used to playtest mechanics before real assets exist.
08 / 13
MEMORY
Your AI Forgets You Every Session.
The shared pain nobody had solved: every session starts from zero.
The cost
Re-onboarding, forever.
The model forgets across sessions, so context has to be rebuilt every time. One discipline in the room: treat it as an always-wrong junior intern and spend up to a week retraining it session by session before it is allowed near the IP.
A week of retraining is a real cost. It was described as worth paying, which tells you how high the downside was judged to be.
The fixes floated
Give it an outboard memory.
External summary documents it reads on every start. A searchable conversation database so past sessions are retrievable. Narrow-context sub-agents that only ever need to know one small thing.
All three are the same admission: memory is your engineering problem, not the vendor's.
Where it landed
Nobody expects the model to remember. The teams getting value are the ones who built the memory around it.
09 / 13
WHERE IT CLEARLY FAILS
AI Cannot Read Market Psychology.
From the trading side-projects in the room: a clean, agreed limit.
Where it holds up
Long-term defensive strategy.
Patient, rules-based, unemotional work is a good fit. It writes the backtesting scripts, checks the logic and never gets bored halfway through.
Note what the builders deliberately kept it out of: the actual decision.
Where it breaks
Anything short-term and human.
It cannot read market psychology, and short horizons are almost entirely psychology. The patterns swapped in the room were behavioural, not technical: Monday morning capital shifting from Asia to Europe, the New York open, Friday sell-offs into the weekend.
A model can learn the pattern. It still cannot feel the room that creates it.
Why it matters beyond trading
The same boundary shows up in games: AI is strong on the systematic layer and weak wherever the answer depends on reading people.
10 / 13
THE SYNTHESIS
Control, Not Capability, Is The Constraint.
Nobody's real problem was whether AI could do the task. It was whether they could trust it and steer it.
The operating reality
Everyone capable was fighting the same battle: keeping AI on-rails.
The bottleneck moved from writing code to reviewing it.
AI belongs in the boring middle: alignment, placeholders, verification and drafts.
The judgement call
Data informs, taste decides. Over-indexing on data leads to trend-chasing and homogeneity.
Small empowered teams beat big ones, and AI widens the gap.
One decision-maker outruns a committee.
The adoption task
Adoption is a culture and economics problem, not a tech one.
The blockers were resistance, budget and trust.
Adapt, do not cut. Fold AI into the pipeline you already run.
11 / 13
FROM THE HOST
One Key. Every Frontier Model.
A note on the tax-and-dividend debate: the cheapest model is often the expensive one once correction time is counted. TokenHub exists so that swapping to a better model is a config change, not a procurement cycle.
What it is
One gateway to the leading open models.
27 models from 5 frontier labs behind a single contract, a single key and a single bill, through Tencent Cloud. OpenAI-compatible, so most stacks move with a base URL change.
Models in play include GLM, Kimi, DeepSeek, MiniMax and Hunyuan.
Why it came up
It lowers the cost of changing your mind.
Vendor list price with no gateway markup, published cached-token rates, and a 99.5% monthly availability SLA with service credits. Run a four-lab evaluation in an afternoon on one key and one cost report.
No lock-in by design: leaving is the same base URL change as arriving.
GLM-5.2
Z.ai
$1.40
out $4.40 · cached $0.26
Kimi K3
Moonshot AI
$3.00
out $15.00 · cached $0.30
DeepSeek-V4-Flash
DeepSeek
$0.14
out $0.28 · cached $0.028
Input price per million tokens, USD, Tencent Cloud list price, Singapore region, published 5 August 2026. Intro deck: tokenhub.tencentcloud-eu.com
12 / 13
YOUR TURN
Which one is
bigger for you?
Every team feels the tax. Not every team can name the dividend. That is the question worth taking back to your studio.
LondonHelsinkiParisNext up
AI First · a demo-first community for AI-native builders in games and creative work
13 / 13