Jack Williams on iTero, GIANTX, and the way forward for AI teaching in esports

iTero, the AI-powered teaching app from GIANTX (previously XL Esports) lately hit the 1,000,000 obtain milestone. The milestone comes after a powerful interval of development following the 2024 acquisition by the LEC group.

The story of the corporate begins rather less glamorously, with Jack Williams, a former Deloitte worker discovering himself banned from each UK betting web site for successful too constantly when betting on esports. iTero was born from a go to to Twickenham, and a realisation that the mannequin he’d constructed might assist skilled (and, with a shove from Overwolf, your common League participant, too).

We sat down with founder Jack Williams to speak by means of iTero‘s journey from a betting mannequin capitalizing on poor pricing to being an LEC scouting machine. We coated every little thing from Riot’s rulesto the place AI teaching is headed subsequent, and when it turns into AI-assisted dishonest.

ESI: Congratulations on hitting the 1,000,000 obtain milestone with iTero. Might you give some background in your success and the way iTero happened? 

Jack Williams: Let’s begin from the highest — I’ll inform the total story. Truthfully, what pursuits me most is the AI-and-esports angle. A number of that work has fed into the buyer product, which has executed very well, and it’s a genuinely helpful inner software — however by itself, “an app that helps individuals get higher on the sport” solely makes for a lot attention-grabbing studying. What individuals truly wish to know is find out how to beat a high group on the weekend and make it to Worlds.

I used to be beforehand in skilled providers as an AI engineer — earlier than “AI engineer” was actually a job title, so on the time, it was extra of a knowledge scientist function. First at HSBC, then at Deloitte. I’ve a Grasp’s Diploma in AI, so I’m pretty technical on that facet. Throughout COVID, whereas at Deloitte, I used to be betting on esports video games — not as a result of I used to be a degenerate gambler, however as a result of I felt I had an edge. 

I observed the betting odds weren’t shifting as soon as the draft was locked in, which struck me as odd, as a result of everybody within the scene is aware of the draft issues — it’s not an ideal predictor of who wins, nevertheless it not less than tells you whether or not a group bought a robust decide within the first rotation, or bought counter-picked.

So I began constructing AI fashions that analyzed the sport state on the level of the draft and predicted the winner from that. I made a bit of cash on the facet doing it — till my betting accounts bought banned. Apparently, within the UK, bookmakers can simply minimize you off in the event you begin successful constantly, which I assumed was a bit unfair. 

However I now had this software and no actual use for it, so I reached out to Kieran Holmes-Derby (Co-Founding father of Excel Esports) and principally stated, “I’ve constructed this factor, it really works, and I’ve been banned from each betting web site. Might I are available in and present your esports group, simply to see if it’s helpful?”

I went to see him at what was then their base in Twickenham Stadium — a bit surreal, strolling as much as Twickenham with the England rugby group coaching on the pitch and the esports group coaching contained in the stadium. I got here away from that considering there needed to be one thing I might construct to assist gamers draft higher and win extra video games. So I resigned from Deloitte and went all-in on the concept: construct instruments for the group, assist them draft, they win extra, everybody’s pleased.

ESI: How difficult was it to depart a full-time job and launch a standalone product again then? 

Jack: I spent about six months constructing it and three months making an attempt to promote it, earlier than realizing that in 2020, no person might afford — or notably needed — something AI-related. It wasn’t even a modern idea but. Extra importantly, most groups didn’t have competent analysts in any respect; it was largely individuals on Google Sheets making issues up, with no actual grasp of pattern measurement. 

Strolling into that setting and saying, “Right here’s a sophisticated, black-box AI mannequin telling you to select this champion, belief me” — that didn’t land nicely with coaches or gamers. I bought zero contracts and was pretty caught, so I deliberate to return to Deloitte.

Then, at an ESI convention, I used to be approached by Overwolf — I’d gained a pitch competitors that 12 months, so I’d made a little bit of a reputation for myself. Somebody there requested if I’d thought-about constructing the software for shoppers as an alternative {of professional} groups. I hadn’t, however figured it was price a strive. Three months later we had a dwell product. It regarded genuinely horrible, however we went from zero to 10,000 downloads in a short time, and retention was a lot larger than typical apps in that class — most apps on this house see 10–20% retention, with individuals uninstalling nearly instantly. Ours held on: individuals thought it regarded tough however favored what it did for his or her drafting, in order that they saved it.

Overwolf then gave me a non-dilutive grant — no fairness taken — which let me rent a developer, and we spent the subsequent two years constructing it into what it’s at present: an AI companion app that helps gamers progress towards their subsequent rank. After two years we had been at round 100,000–200,000 downloads. Then we had been acquired by GIANTX — initially XL Esports — which introduced issues full circle again to Kieran. That occurred in 2024.

ESI: And post-acquisition, what modified for the corporate, and the product? 

Jack: As soon as I used to be built-in into the group, the main focus grew to become pure development: how will we attain extra gamers and construct a greater product? We shortly discovered actual synergy between an esports group making an attempt to win the LEC and a product serving to the typical solo-queue participant win their subsequent sport — synergy in decision-making, since I now had entry to coaches and gamers who’re genuinely glorious on the sport, which improved the mannequin, and synergy in advertising, because it’s a lot simpler to advertise a software like this if you’re backed by an esports group.

Quick-forward two-plus years, and we’ve crossed one million downloads. We had been at roughly 500,000 final 12 months, and round 200,000 the 12 months earlier than — so very fast development since becoming a member of GIANTX. Customers have grown roughly 300% over the identical interval — downloads have doubled, lively customers have tripled. We’re in an ideal place with the product proper now; we expect it’s one of the best available on the market by a long way.

We’re additionally enthusiastic about what the group can do extra broadly. GIANTX is understood for locating rookies — we don’t have the most important funds within the league, however we constantly end round fourth. A part of that comes from the statistical fashions we use to establish new gamers — I wouldn’t essentially name them AI, however they’re superior. That’s helped us overperform, and whereas many of the credit score goes to the coaches and gamers, some goes to the methods we’ve constructed by making use of the ideas from our solo-queue AI instruments to skilled scouting.

ESI: Nice stuff. You talked about it’s been a mad 12 months of development — what do you suppose actually drove that? Was it merely extra funding within the product?

Jack: There are some things, and I don’t wish to miss any of the vital ones, however basically a very powerful issue was that product was at all times the precedence — at each stage. We by no means determined to only fill the app with advertisements for fast income, or run low-cost, spammy advert campaigns to inflate obtain numbers with no actual substance. The one factor I cared about was retention. Due to that, most of our funding went into enhancing the product and the retention price, to the purpose the place most individuals who obtain the app are nonetheless utilizing it every week later — I don’t suppose every other app within the house can declare that over half their customers stick round after every week, although there’s no public knowledge to match in opposition to.

So: product and retention first. Second, GIANTX gave us an actual benefit as a Spanish group with sturdy ties to Spain. Virtually all of our development has come from Spanish-speaking Latin America and mainland Spain. Once we joined the group, roughly 20% of our customers had been Spanish-speaking — that’s now round 70%. If I listed our high 5 international locations by customers, each single one can be primarily Spanish-speaking — Spain, and I consider Argentina, Chile, Mexico, and another in Latin America.

ESI: I feel you’re proper…

Jack: What we had been in a position to do was lean on our personal group: an organization filled with Spanish audio system with real Spanish-language networks, to develop organically. We reached out to native content material creators, often pretty small ones, and labored with them on a private stage: “Right here’s the software, strive it your self, inform us truthfully what you suppose. We’re not going to ask you to oversell it. For those who genuinely don’t prefer it, let’s not work collectively. But when there are elements you do like, make a video about it.” That form of natural product placement performs much better than a scripted advert that simply hits advertising speaking factors. If creators genuinely just like the product, they speak about it nicely.

Third, GIANTX has its personal inner instructional content material channels — rebranded a few occasions. I don’t know the precise figures, however we’re speaking a whole lot of thousands and thousands of impressions yearly throughout these movies — ideas, tips, “how do you have to play this,” that form of content material. That content material naturally lends itself to selling the product: “Right here’s the optimum construct this patch — if you wish to know the optimum construct at any time, obtain the app.” Very pure placement by means of channels we already personal.

ESI: I recall — and I is perhaps paraphrasing the precise wording — that Riot’s coverage bars third-party instruments from creating an unfair benefit gamers wouldn’t in any other case have. Have you ever ever had that query raised, or needed to pull again a function as a result of Riot stated no?

Jack: With Riot, we’re very cautious with them. We’re most likely essentially the most conservative of the third-party apps: if Riot tells us to not do one thing, or the wording is ambiguous, we merely don’t do it, as a result of staying on the suitable facet of Riot issues to us. Given our esports connection, we’re deeply embedded within the Riot ecosystem, and we don’t wish to jeopardize that. We additionally need gamers to belief the app — you don’t need somebody downloading it questioning if it’s functionally just like an aimbot. So we keep away from any gray-area options solely; we solely construct what’s clearly permitted.

Riot’s personal wording isn’t particularly exact — their third-party instruments coverage says instruments can’t present a “aggressive benefit,” however that’s a fuzzy customary. Technically, sorting a stats web site by win price and selecting the highest champion can be utilizing obtainable knowledge to make one of the best resolution — that’s a aggressive benefit too, by that logic. In my opinion, what would truly be unfair is surfacing data that isn’t accessible by means of the sport’s personal UI — pulling from the sport’s underlying code, capturing occasions that aren’t meant to be seen, and displaying them in methods Riot doesn’t sanction.

Typically Riot’s selections are a bit arbitrary, however we simply comply with them. For instance, summoner-spell timers — a easy UI factor you’d click on when somebody used Flash, exhibiting a countdown till it was again up — had been pretty customary throughout instruments. Riot determined they now not needed them within the ecosystem, so we eliminated ours. Usually we keep conservative, and if Riot ever asks us to alter one thing, we do it instantly.

ESI: How do you see this altering as tech continues to develop and develop? 

What’s going to get extra attention-grabbing is when it turns into more durable for Riot to implement this sort of management. Proper now, they management me largely by means of API entry — if I upset Riot, they pull my API entry, and I now not have an app. That’s a major deterrent.

However instruments that run superior AI fashions straight on a participant’s monitor, producing their very own knowledge by studying in-game visuals slightly than utilizing the API, are a lot more durable to manage. That’s a genuinely regarding course. Perhaps much less so in League, extra so in FPS titles. 

You could possibly find yourself with one thing operating invisibly, not on the pc however successfully studying the monitor, offering an unfair benefit that’s very arduous to detect. For instance, a high-end monitor may be capable to establish precisely the place somebody’s head is on display, or feed directions a participant wouldn’t in any other case have. That’s a worrying house for aggressive integrity. However broadly, we keep extra conservative than we strictly have to be, particularly to remain in Riot’s good graces.

ESI: Do you ever fear that somebody may simply copy your function, given the shortage of IP safety? 

UI is trivial to duplicate now with AI coding instruments. Somebody can screenshot a UI, feed it to an AI mannequin, and ask it to recreate the interface. What’s genuinely troublesome , and even near not possible at present, is recreating the predictive fashions behind it. These outcomes come from a variety of coaching, a variety of knowledge, and a variety of time. Which will turn into simpler for AI to duplicate ultimately, and doubtless earlier than individuals count on, however as of at present, you may copy our UI — you may’t copy the underlying intelligence.

ESI: How does it work being embedded inside an esports group? Do you continue to supply providers to different skilled groups, or does belief turn into a difficulty given your GIANTX affiliation?

Jack: No — we work solely inside GIANTX, together with our LEC group and our Spanish league (LVP) group, and would lengthen to every other titles we compete in going ahead. We’re internal-only. There’s actual demand on the market. Groups with modest budgets would love a software like this to seek out undervalued gamers, however for now we haven’t opened it as much as different groups, purely for aggressive integrity causes. Any group with the funds to pay for one thing like this can be a group we’re more likely to face sooner or later, so for now the reply is not any, although that might change in future.

ESI: Does your mannequin attempt to account for much less quantifiable elements — participant tilt, greed, tendencies that don’t present up cleanly within the last match knowledge? Do you break issues down by early-game phases, and the way deep does the modelling go on the aggressive stage?

Jack: Due to precisely the problems you’re describing, we have a tendency to stay to solo queue — it’s a lot better suited to AI predictive modelling. There’s much more knowledge, and also you keep away from the noise of, say, one Fnatic-vs-G2 consequence on one patch with two particular drafts. There’s little or no to study from a single knowledge level like that. In solo queue, you may observe the identical or comparable draft repeated at excessive elo many occasions over and draw actual conclusions.

Most of what Itero does day-to-day is solo queue evaluation slightly than pro-play prediction. You’re proper that the variety of variables concerned in pro-play prediction is gigantic — even with good knowledge, you’re most likely capped round 80–85% accuracy. 

And even a modest edge, say a 15% likelihood of an upset, will nonetheless play out not less than as soon as throughout a cut up if a group performs ten video games. 15% sounds low, however throughout sufficient matches it nonetheless exhibits up. There are all types of tales, not simply in esports however in soccer too, of individuals discovering spurious correlations. 

This will take the form of monitoring a participant’s Instagram exercise, relationship standing, no matter. Chasing an “edge” usually is absolutely simply noise. Anybody could make a knowledge level say nearly something in the event that they ignore causation versus correlation.

ESI: Has your mannequin ever contradicted established draft principle — going in opposition to what’s thought-about “meta” — and ended up shifting the meta itself? Or does it typically reinforce what’s already accepted?

Jack: On the professional facet, I don’t wish to overstate our affect. These selections actually sit with the coaches. Any notable calls or improvements are all the way down to the teaching employees and gamers; I wouldn’t declare we’ve reshaped the LEC meta. The place you most likely would discover a sample is in solo queue. As extra individuals use the software, extra gamers are taking objectively favorable matchups. 

We’re now at a stage of market penetration the place it’s seen within the knowledge: even at excessive elo, the place utilization is comparable throughout the elo spectrum, extra drafts are being optimally picked than earlier than the software existed.

That will get attention-grabbing at a deeper stage, although. If drafting optimally turns into so widespread, by means of us or competing instruments, that everybody can predict “if I decide X, they’ll decide Y,” gamers begin on the lookout for a 3rd transfer: flex picks, or off-meta picks. 

Say you’re a Vayne high participant, and each time you decide Vayne high, you understand the opponent will decide a tough counter like Teemo, you may begin swapping roles mid-lobby extra usually: “I’ve been counter-picked once more, let’s change it up — I’ll go Vayne ADC, another person take high.” So there’s potential for that form of disruption.

On the professional facet particularly, I feel our scouting method has been extra attention-grabbing. What we’ve executed is establish gamers who’re basically excellent on the sport mechanically, after which construct teaching help round them. 

That is the alternative of the previous method of selecting gamers purely as a result of they’ve been within the scene for years, know find out how to deal with scrims, dwell the “professional participant” life-style, or can address stage stress. That method limits you to a small, recycled pool of “identified” gamers. 

We weren’t fascinated with that. We search for somebody basically sturdy on the sport, which doesn’t essentially imply the highest-ranked solo queue participant. Ranked talent doesn’t routinely translate to performing nicely on stage. We don’t simply rank by LP and decide the highest identify. However we’ve constructed a system that identifies gamers more likely to be basically sturdy performers on stage, carry them in no matter background, and construct the supporting methods round them.

Jackie is our greatest instance. He wasn’t coming from a extremely aggressive LFL or EU Masters group. We discovered him by means of the info, introduced him in, constructed the methods round him, and he’s now genuinely our MVP in that function, on account of a fundamentals-first method slightly than conventional esports scouting.

ESI: From a participant expertise angle — pre-game, in-game, post-game — which stage do you suppose issues most for somebody making an attempt to enhance? Do individuals principally use it as a decide overlay and transfer on, or are individuals meaningfully partaking with post-game experiences? Do you observe utilization patterns to see which behaviors correlate with essentially the most enchancment?

Jack: It’s an odd one, as a result of ideally you’d measure it by win price, however the software is basically designed to push everybody again towards a 50% win price over time. Play higher with the software, and also you climb elo till you stabilize once more — that’s only a statistical actuality. 

So beneficial properties from the software are typically short-term by nature, however actual. What we do know: as soon as somebody begins utilizing the software, their win price rises as a baseline. And in the event you comply with the champion suggestions carefully, your win price rises fairly considerably within the quick time period. 

We’re speaking roughly a 6% enhance. That sounds modest, however sustained over time, that’s sufficient to climb from, say, Gold to Platinum simply by following the highest suggestion constantly.

The primary friction isn’t actually a difficulty, extra a alternative gamers make: individuals don’t love at all times enjoying the very same really useful champion. For those who’re first or second decide, the highest advice is usually just like what was instructed the sport earlier than, since not a lot has modified. Perhaps you understand one teammate and a few enemy picks, however except it straight impacts your lane matchup, the advice doesn’t shift dramatically. So individuals search for selection. This implies turning down the “prioritize meta decide price” weighting, as an example, to floor extra off-meta choices.

The clearest instance is our champion-mastery weighting. By default it elements in how a lot expertise you will have on a champion, which is perfect for successful. Left on the default setting, you’ll win extra. Virtually everybody turns it all the way down to zero, like, “I don’t care about my expertise, simply inform me one of the best decide,” then first-times a random champion in ranked and loses. 

That’s superb, truthfully; the sport isn’t just for individuals making an attempt to optimize their climb. Some individuals simply wish to have enjoyable and use it to find new champions. Essentially, it’s arduous to assert the software “does” a sure win price, as a result of most individuals voluntarily set it to be much less optimum, as a result of that’s extra pleasurable.

ESI: The place do you see AI teaching heading throughout gaming and esports extra broadly? Do you suppose it’ll turn into customary, one thing publishers construct themselves, and will it assist stage the worldwide aggressive panorama?

Jack: That’s a loaded query, so let me attempt to unpack it. First: I feel it’s unambiguously good for sport engagement typically, separate from esports solely. If League of Legends existed with no esports scene in any respect, would instruments like this nonetheless make the sport higher? I feel sure. And whether or not that comes from third events like us or from publishers constructing it themselves, I feel this class will continue to grow and enhancing, hopefully in a wholesome course.

My favourite reference level is chess.com. It’s excellent at leaving the precise sport solely as much as you — no transfer strategies, no dwell teaching, no win-probability readout when you play. However the pre-game coaching and particularly the post-game evaluation are glorious. That’s actually the course I’d wish to see video games go: making it clearer why you misplaced, so it feels much less arbitrarily irritating. 

The toughest a part of League, in the event you’re enjoying a few video games a day at a 51% win price — which is genuinely good, and can get you climbing ultimately, is that every particular person sport nonetheless appears like near a coin flip, which is demoralizing within the second.

What these instruments allow you to do is give gamers one thing to work on that isn’t merely “win or lose.” We’ve a macro coach that opinions your latest video games and tells you, for instance, “over your final fifty video games, right here’s the error you make most frequently” — so you may direct your observe time intentionally. 

Say the sample is: you win lane, however constantly lose expertise between the 14- and 21-minute mark. You possibly can go into your subsequent sport centered particularly on enhancing that quantity, and that provides you a way of management over your personal enchancment, impartial of whether or not the opposite 9 gamers in that sport trigger you to win or lose. You possibly can lose the sport and nonetheless be ok with it, as a result of the particular factor you had been engaged on — not throwing away your early lead — truly improved. That’s basically higher for the sport and the participant.

For esports particularly, will it shut the hole with Asia? Most likely not, largely as a result of it comes all the way down to economics and funding. In case your regional market doesn’t generate sufficient income, you may’t reinvest in enchancment, and it turns into very troublesome to catch up. 

I gained’t fake to have the reply for fixing European esports economics. Everybody is aware of it’s a tricky setting. Essentially, these instruments price cash, the identical nearly as good coaches, good gamers, good amenities. Even issues like a fitness center and a efficiency coach connected to a group present measurable profit. We’ve seen that within the knowledge, and the identical with sleep monitoring and participant wellbeing work. Can European groups broadly afford all of that? No. Identical logic applies to AI tooling. For those who can’t afford to construct it, you may’t profit from it as a lot.

Will it enhance everybody over time? Sure. We’ve seen this in soccer, the NFL, baseball, basketball, it doesn’t matter the game. AI helps groups enhance. But it surely helps everybody enhance roughly proportionally, so the relative hole doesn’t essentially shut. We’ll most likely nonetheless be shedding to Korea in ten years. We would have AI mannequin by then, however they’ll probably have a greater one.

ESI: So, it appears like in-game teaching is one thing you wish to swerve? 

Jack: I touched on it with the chess instance. Not everybody out there shares that view, so it’ll be attention-grabbing to see how Riot handles it as instruments get extra succesful.

That’s when hen in-game instruments begin veering towards AI-assisted dishonest slightly than professional teaching. I’ve seen one instance: a “phantom marker” that exhibits an icon on the map on the final location a champion was seen, even after they’re now not seen. That is successfully remembering positioning data for you. That’s nicely over the road, in my opinion. Virtually every little thing past the essential stat-comparison options we have already got — like “how does your farm evaluate to the typical participant at your elo” — begins to cross into that territory.

There was one other one which used voice prompts — actually telling you, “go do Dragon now,” “go do Baron now.” That felt like a transparent step too far, particularly once they used recognizable esports personalities’ voices for it. Enjoyable as a novelty, however nicely previous the road.

ESI: With enchancment focuses such because the 14-to-21-minute XP monitoring instance, do gamers ever tunnel on that particular metric? Like, enhancing their very own XP quantity whereas their group falls aside in each different lane. Is there a threat that focusing so closely on particular person efficiency metrics undermines general group play and win price?

Jack: I nervous about that too, truthfully. What we’ve truly seen is near the alternative. Gamers who’ve performed hundreds of video games develop very fastened habits, and the toughest half is breaking that routine in any respect. 

If we are able to quickly flip a devoted farmer into extra of a ganker, even for a handful of video games, I feel that’s internet constructive. Even when they ultimately revert to their previous play type, as a result of they’ve picked up a brand new talent alongside the best way. 

Identical logic as a help participant in CS out of the blue enjoying entry for a stretch: even when they return to help afterward, they’ve gained one thing from the expertise, and so they’ll probably perceive that play type higher when going through it too. It’s superb if a participant over-indexes on one enchancment space for a few video games and dips elsewhere, that’s a part of studying to stability it. Once they revert to their standard type, they carry that additional talent with them.