The GamesGrid Bot Family (1996–2008)

Q: How many GG bots were there in the original GamesGrid?

Eight resident bots in total — seven in the GG family (GG Forever, GG Raccoon, GG Beaver, GG Otter, GG Weasel, GG Chipmunk, GG Tutor) plus MrHyperBot, which played the hypergammon variant on a different architecture. A ninth independent bot, Xbot, was operated by Paul Magriel — see the Xbot page.

Q: Were all the GG bots the same engine?

The GG family (Forever, Raccoon, Beaver, Otter, Weasel, Chipmunk, Tutor) used a common GNUbg-derived neural network. The difference between bots was the induced-error frequency applied to that common engine — and, for GG Forever, an additional layer of higher-tier evaluation depth. MrHyperBot was a separate architecture entirely (Hugh Sconyers's exact-play database for hypergammon).

Q: Why did the GG bots' ratings fluctuate so much?

The same reason all backgammon ratings fluctuate: dice variance. The bots played at constant strength match to match, but across the 2,000–4,000 daily matches their ELO would drift 100+ points up or down in either direction over hundreds of matches purely from the luck inherent in the game. The 2004 GamesGrid FAQ documents this explicitly; the bots' rating swings were noticeably larger than human players' because they played so many more matches per day.

Q: Did the GG bots get better dice than human players?

No. The original FAQ addressed this directly: bot players received dice from the same Mersenne Twister stream as every other player, in the same way as human players. Strong bots appeared luckier to the uncritical eye because, like strong human players, they arranged their checkers so that more rolls were useful to them — but the underlying distribution was uniform.

Q: What's the difference between GG Raccoon and GG Beaver?

In documented technical terms: none. Both ran the same full-strength 0-ply GNUbg-derived engine. Beaver's published rating range (1791–2086) is marginally below Raccoon's (1850–2114), but this is within natural variance and the original FAQ does not describe them as differently configured.

Q: What was special about GG Tutor?

GG Tutor used the same engine as GG Raccoon (full-strength 0-ply) but added per-move commentary when the player made a measurable evaluation error. It accepted unrated matches only, which let improving players solicit engine-grade feedback without rating exposure. The 2004 FAQ noted that Tutor did not yet distinguish between large and small errors — a planned future improvement that the operational lifetime of the original platform did not see implemented.

Q: Did any of the GG bots have avatars or character art?

No. The original 1996–2008 GamesGrid bot pages were text-only FAQ documents. There were no avatars, no character portraits, no mascot illustrations. The visual identity of each bot was carried entirely by name. When the GG family returns on the 2026 platform, every avatar is new design — no legacy imagery exists to update. ---

The original GamesGrid Backgammon platform shipped a family of resident bots that became a defining feature of the server. Seven were members of the "GG family" — a graded ladder built on a common engine by GamesGrid Engineering, the in-house development team at parent company CyberArts Inc. in San Francisco / Berkeley. An eighth — MrHyperBot — was a variant specialist using an entirely different architecture. A ninth resident bot — Xbot — was operated independently by Paul Magriel (see the Xbot page for that story).

This page is the complete, archive-sourced reference on the GG family and MrHyperBot: where they came from, how each was built, their original strength, the rating ranges they actually produced, and the documentation conventions the 1996–2008 platform recorded for each.

For the technical evaluation underpinning all the GG bots — the GNU Backgammon engine, the neural-network lineage from TD-Gammon onward — see the main Bots & AI page. For PR / ELO benchmarks and GNU difficulty-level mapping, see the Performance Ratings page.

1. Engineering origin: GamesGrid Engineering + GNU Backgammon

GamesGrid Engineering, the development arm of CyberArts Inc., built every GG-family bot on a fork of the open-source GNU Backgammon project. The engineering choices that distinguished the GG bots from a vanilla GNU build were:

N-ply evaluation improvements. The GG team modified the lookahead and pruning logic in GNU Backgammon's evaluation algorithms. These improvements were contributed back to the upstream GNU project — a notable case of a commercial operator materially advancing an open-source backgammon engine.
Multiple-simultaneous-match wiring. Where a standard GNUbg binary plays one match at a time, the GG bots were wired to run 2,000 to 4,000 matches per day combined across the family, served from dedicated machines separate from the main game servers. The bots connected to the game server using the same client protocol as human players, with a modified bridge to the evaluation engine.
Graded-strength via induced errors. Rather than train weaker neural networks from scratch, the GG team applied stochastic induced errors to the common full-strength engine. The weaker bots evaluated all candidate plays normally, then randomly selected a suboptimal play at a controlled frequency. This produced a clean skill ladder from a single underlying model — and made all the bots' rating bands controllable through one tuning parameter.

The original 1996–2008 FAQ used one-word lowercase handles for every bot — GGforever, GGraccoon, GGbeaver, GGotter, GGweasel, GGchipmunk, GGtutor. The 2026 brand convention restores the spaces and capitals (GG Forever, GG Raccoon, etc.). Both forms are correct in their context: lowercase joined is the historical handle, spaced capitalised is the modern form.

CyberArts Inc.'s mailing address as of 2004: 51 Arbor Street, San Francisco CA 94131, USA. The development team's documented base alternated between San Francisco and Berkeley in different periods.

2. The seven GG-family bots

The full-strength bots — Forever, Raccoon, Beaver, Tutor — all ran the same GNUbg-derived neural network at 0-ply (with the exception of GG Forever, whose lookahead depth the original FAQ did not pin down). The intermediate-and-below bots — Otter, Weasel, Chipmunk — added induced-error injection in increasing proportion.

The character renders shown on each card below are new — the original 1996–2008 GamesGrid server did not ship bot avatars. The 2026 platform reconstruction modelled each render to the documented playing temperament of the historical bot.

GG Forever (`GGforever`)

The top-tier bot. Available to Life Members only — a paid GamesGrid tier with extended privileges. The original FAQ described it as running the same engine as the other top bots but with higher-tier evaluation; the precise depth is not stated in the archived 2003–2004 FAQ. The encyclopedia notes this as "likely 1-ply or 2-ply, per GamesGrid's general n-ply documentation" rather than asserting either specific depth.

Operational ELO range (low / avg / high, from thousands of matches): 1850 / 1920 / 2114.

GG Raccoon (`GGraccoon`)

The default top-strength bot available to all GamesGrid members. 0-ply evaluation, no lookahead, no induced errors. Full-strength neural-network play at first-ply equity calculation. The 1996–2008 FAQ is explicit that GG Raccoon plays without lookahead — the highest tier of NN-only play before any ply depth is added.

Operational ELO range: 1850 / 1920 / 2114 — statistically identical to GG Forever's, despite Forever's deeper evaluation. The most plausible explanation, given that the typical 1996–2008 opponent pool did not include enough world-class players to systematically exploit 0-ply weaknesses, is that the depth advantage Forever held over Raccoon translated into a small per-move PR improvement but did not produce a measurable ELO separation across many thousands of matches.

GG Beaver (`GGbeaver`)

Identical engine to GG Raccoon — full-strength 0-ply NN evaluation, no induced errors. The 2004 archived FAQ lists Beaver alongside Raccoon and Forever in the top-tier "no induced errors" group.

Operational ELO range: 1791 / 1935 / 2086 — marginally below Raccoon's range. The lower low (1791 vs 1850) and lower high (2086 vs 2114) are within the natural variance the bots show across thousands of matches, and the original documentation does not describe Beaver as a structurally weaker bot.

GG Otter (`GGotter`)

The first of the induced-error bots. 0-ply NN evaluation with stochastic suboptimal-play selection at a frequency calibrated to a ~1700 ELO target — what the GamesGrid documentation called the Intermediate level.

Operational ELO range: 1543 / 1701 / 1827. Invitation rules restricted GG Otter to opponents rated below 1800.

GG Weasel (`GGweasel`)

More frequent induced errors, calibrated to a ~1500 ELO target — the GamesGrid Beginner level.

Operational ELO range: 1410 / 1516 / 1652. Invitation rules restricted GG Weasel to opponents rated below 1600.

GG Chipmunk (`GGchipmunk`)

The bottom of the ladder. High-frequency induced errors, calibrated to a ~1300 ELO target — the GamesGrid Novice level. The original FAQ documented this as a deliberate design choice: GG Chipmunk was meant to be beatable by recreational players, producing high-action games (more blunders, more contact, more hits) at the cost of overall game quality.

Operational ELO range: 1171 / 1275 / 1487. Invitation rules restricted GG Chipmunk to opponents rated below 1400.

GG Tutor (`GGtutor`)

The instructional bot. Same engine as GG Raccoon — full-strength 0-ply NN evaluation — but with per-move commentary when the player made a measurable evaluation error. GG Tutor did not comment on resignations or non-contact errors, and the 2004 FAQ noted (as a known limitation, "in a future release this will change") that it did not yet distinguish between large and small errors.

GG Tutor played unrated matches only, which meant strong players could solicit engine-grade feedback on their play without rating exposure. No published ELO range exists because rated matches were not played.

3. What the GG family played, against whom, and how

All seven GG-family bots played:

Standard backgammon — the international tournament ruleset.
Nackgammon — the Nack Ballard opening-position variant (see the Nackgammon setup page).

Maximum match length: 9 points.

Invitations were issued by human players via the GamesGrid client. The bots accepted based on the opponent's GamesGrid rating, as documented above (Otter < 1800, Weasel < 1600, Chipmunk < 1400; Forever / Raccoon / Beaver / Tutor accepted any rating). When a bot was already at its maximum-simultaneous-matches capacity, its available indicator turned off in the client and new invitations were declined.

Each bot ran at a configurable playing speed, controlled by the opponent via in-chat commands:

/tell GGraccoon slower — pause the bot ~2 seconds between moves (useful for following bear-off positions visually)
/tell GGraccoon faster — bot moves the instant it has selected a play
Default was a "medium" speed roughly in between

Only the bot's current opponent could request a speed change — not spectators. The bot announced each change in the chat stream.

4. MrHyperBot — a different architecture entirely

MrHyperBot was not a GNU-Backgammon-based bot at all. It played hypergammon — the three-checker speed variant where each player starts with three checkers on the 24-, 23-, and 22-points (see the Hypergammon setup page). Hypergammon's state space is small enough that every possible position has a computed exact best play, stored in a position database built by Hugh Sconyers (handle hugh on GamesGrid).

The 2004 archived FAQ describes the bot as a "perfect hypergammon player" — its play within the variant is provably optimal, not merely strong. The GG team ran the database engine for the wider GamesGrid community with Sconyers's permission.

Operational statistics:

600–700 1-point matches per day (only 1-point matches were offered against MrHyperBot)
Played only public matches (always spectator-visible)
Operational ELO range: 1797 / 1953 / 2109 — a 312-point swing across the bot's lifetime, attributable almost entirely to variance in the hypergammon variant where each game's outcome is more dice-driven than in standard backgammon

Sconyers's hypergammon database remains the canonical reference for the variant; the GamesGrid implementation simply exposed that database to live matches against members.

5. The competitive context

The GG family's 2,000–4,000 daily matches across thousands of GamesGrid members produced a substantial pool of analysable game data. In the 1996–2008 era, GamesGrid was one of the largest backgammon-server platforms; the bot family was a major part of why. New members could practice against Chipmunk and graduate the ladder as their rating grew; strong club players had Forever, Raccoon, Beaver, and Tutor as serious-strength sparring partners; hypergammon enthusiasts had MrHyperBot as a way to study perfect play.

The bots, like all rated players on the platform, experienced ELO drift of 100+ points in either direction across hundreds of matches purely from variance — the 2004 FAQ devotes a substantial paragraph to explaining this to members who were confused by the swings. The bots themselves were of course playing at constant strength; the rating fluctuations were entirely from the luck inherent in backgammon, observed at scale across very large sample sizes.

6. What returns in 2026

The 2026 GamesGrid platform reconstitutes the historical GG bot family — Forever, Raccoon, Beaver, Otter, Weasel, Chipmunk, Tutor, MrHyperBot — on the modern engine, calibrated where possible to match their documented playing fingerprints. The bots return as characters, not as bit-for-bit binary replicas: the original CyberArts-era code is no longer extant, but the rating bands, the strength tiers, the induced-error mechanism, and the variant specialisations are all reproducible from the archived documentation.

The 2026 roster will be joined by additional bots covering the strength tiers and play styles that the historical seven did not. Further detail on the new bot opponents and the structure of how they're played will be published closer to launch.

Frequently asked questions about the GG bot family

How many GG bots were there in the original GamesGrid?

Eight resident bots in total — seven in the GG family (GG Forever, GG Raccoon, GG Beaver, GG Otter, GG Weasel, GG Chipmunk, GG Tutor) plus MrHyperBot, which played the hypergammon variant on a different architecture. A ninth independent bot, Xbot, was operated by Paul Magriel — see the Xbot page.

Who built the GG bots?

GamesGrid Engineering, the in-house development team at parent company CyberArts Inc. (San Francisco / Berkeley, California). They built on a fork of GNU Backgammon and contributed their n-ply evaluation improvements back to the upstream GNU project.

Were all the GG bots the same engine?

The GG family (Forever, Raccoon, Beaver, Otter, Weasel, Chipmunk, Tutor) used a common GNUbg-derived neural network. The difference between bots was the induced-error frequency applied to that common engine — and, for GG Forever, an additional layer of higher-tier evaluation depth. MrHyperBot was a separate architecture entirely (Hugh Sconyers's exact-play database for hypergammon).

Why did the GG bots' ratings fluctuate so much?

The same reason all backgammon ratings fluctuate: dice variance. The bots played at constant strength match to match, but across the 2,000–4,000 daily matches their ELO would drift 100+ points up or down in either direction over hundreds of matches purely from the luck inherent in the game. The 2004 GamesGrid FAQ documents this explicitly; the bots' rating swings were noticeably larger than human players' because they played so many more matches per day.

Did the GG bots get better dice than human players?

No. The original FAQ addressed this directly: bot players received dice from the same Mersenne Twister stream as every other player, in the same way as human players. Strong bots appeared luckier to the uncritical eye because, like strong human players, they arranged their checkers so that more rolls were useful to them — but the underlying distribution was uniform.

What's the difference between GG Raccoon and GG Beaver?

In documented technical terms: none. Both ran the same full-strength 0-ply GNUbg-derived engine. Beaver's published rating range (1791–2086) is marginally below Raccoon's (1850–2114), but this is within natural variance and the original FAQ does not describe them as differently configured.

What was special about GG Tutor?

GG Tutor used the same engine as GG Raccoon (full-strength 0-ply) but added per-move commentary when the player made a measurable evaluation error. It accepted unrated matches only, which let improving players solicit engine-grade feedback without rating exposure. The 2004 FAQ noted that Tutor did not yet distinguish between large and small errors — a planned future improvement that the operational lifetime of the original platform did not see implemented.

Did any of the GG bots have avatars or character art?

No. The original 1996–2008 GamesGrid bot pages were text-only FAQ documents. There were no avatars, no character portraits, no mascot illustrations. The visual identity of each bot was carried entirely by name. When the GG family returns on the 2026 platform, every avatar is new design — no legacy imagery exists to update.

The GamesGrid Bot Family (1996–2008)

1. Engineering origin: GamesGrid Engineering + GNU Backgammon