Technical report · live · not yet updated

The Kid: raising a language model from scratch into a profitable trader

Abstract. The Kid is a 1.38-billion-parameter language model trained from random initialization, in public, toward one objective: to trade profitably. It learns language first, then arithmetic, then markets. Every checkpoint is graded on fixed prompts, kept, and can be queried below. This page reports the model's live training state, its evaluation results, and its progress toward trading.

Model: nanochat GPT, 24 layers
Token: $KID · Robinhood Chain
Objective: Profitable trading
Compute: 8×H100 SXM per spurt
Status: Pre-launch

DAY 0EGG

Subject 001 Stage: Egg · Brain 1

Stage: Egg
Brain: Brain 1
Training step: 0 / 5,568
Tokens seen: 0
Validation loss: n/a
Coherence: n/a
Milestones: 0 / 8
Checkpoints: 0
Age: n/a
Trading: Course 1
Compute spent: $0

Figure 1. Smoothed training loss (nats per token) for the current brain against training step. Numbered markers are milestones from the log in §4. The shaded band is the part of the run not trained yet.

Interactive evaluation

Messages go to a live checkpoint. Before chat fine-tuning the model is a pure next-token predictor: your text becomes a prefix and the model continues it. After fine-tuning it answers in turns. Outputs are shown exactly as generated, with no filtering.

Checkpoint completion temperature 0.8 · top-k 50 · max 48 tokens

systemLoading the latest checkpoint…

Rate limited per visitor. Each message runs on the same checkpoints the grader scores.

Training progress

Pretraining is split into growth spurts. Each spurt resumes the previous checkpoint with its optimizer and data-loader state, so the spurts add up to exactly one uninterrupted run with the same learning-rate schedule. Every spurt ends with a checkpoint, a validation pass and a grading pass.

Figure 2. Validation loss in bits per byte on held-out text at each pretraining checkpoint. Lower means the model is less surprised by unseen text.

**Table 1.** Checkpoints, newest first. Grader metrics are defined in §6.
Checkpoint	Stage	Step	Tokens	Val bpb	Coherence	Variety	Facts	Math	Created
No checkpoints yet.

Developmental stages

Subject 001 is drawn from the model's training state. Its body changes at each stage transition below. A new brain starts from random weights again, at a larger size.

Figure 3. Stages of Subject 001. Filled: reached. Silhouette: not reached yet.

**Table 2.** Stage transitions.
Stage	Triggered by	Training phase	Planned
Egg	Before the first checkpoint	n/a	Pre-launch
Hatchling	Step-0 checkpoint (random initialization) published	Pretraining	Launch hour
Babbler	First pretraining spurt completes	Pretraining, 12 spurts	Days 1–3
Talker	First chat fine-tuning checkpoint	Supervised fine-tuning	Days 4–5
Student	First reinforcement-learning checkpoint	RL on GSM8K	Days 6–7
Teen	Brain 2 begins pretraining	New run, 32 layers	Weeks 2–3
Grad	Brain 3 begins pretraining	New run, 40 layers	Weeks 4–6

Milestone log

A milestone is logged at the first checkpoint where its criterion holds, with the exact output that satisfied it. Criteria are evaluated by the grader after every spurt.

**Table 3.** Milestones reached, in order.
#	Milestone	Checkpoint	Reached	Evidence (verbatim)
No milestones yet.

Milestone criteria

First words that go together: Mean word-pair coherence ≥ 0.25 across the grader's base-model samples.
First sentence: A run of six or more dictionary words ending in . ! or ?, with pair coherence ≥ 0.6 and variety ≥ 0.9 inside the sentence; mean coherence ≥ 0.45 and variety ≥ 0.7 overall.
First fact: At least one of five factual prefixes ("The capital of France is") completed correctly within six words at temperature 0.
First conversation: The first chat fine-tuning checkpoint.
Knows its name: "The Kid" appears in at least two of three answers to name questions.
Knows its ticker: "$KID" appears in at least one of two answers to ticker questions.
First math problem solved: At least one of six grade-school word problems answered correctly.

Checkpoint comparison

The same prompt sent to three checkpoints at once. By default: the first checkpoint, the middle one and the latest.

Evaluation

Two kinds of evaluation. Grader metrics run on every checkpoint and are cheap and noisy. Standard benchmarks run at the end of each training phase and are comparable with published models.

**Table 4.** Grader metrics at the latest checkpoint of each phase, current brain.
Metric	Pretraining	Chat fine-tune	RL	Definition

**Table 5.** Standard benchmarks. The Kid's scores are filled in at the end of each phase.
Benchmark	Measures	GPT-2 (2019)	The Kid, Brain 1
CORE	DCLM composite of in-context learning tasks	0.2565	after pretraining
ARC-Easy	Grade-school science, multiple choice	n/a	after fine-tune
ARC-Challenge	Harder science questions	n/a	after fine-tune
MMLU	57 academic subjects, multiple choice	n/a	after fine-tune
GSM8K	Grade-school math word problems	n/a	after RL
HumanEval	Python function synthesis	n/a	after fine-tune

GPT-2's CORE score is the reference nanochat reports for GPT-2 capability. Chat benchmarks have no comparable GPT-2 figure, so they're marked n/a.

Model and training setup

The model is trained with nanochat, an open-source harness (MIT license). Depth is the single dial: width, heads, training horizon, batch size and learning rates follow from it. Figures below come from nanochat's own sizing code at a tokens-to-parameters ratio of 8.

**Table 6.** Brain specifications. Brain 2 and Brain 3 are the planned successors.
	Brain 1	Brain 2	Brain 3
Layers	24	32	40
Model width (d_model)	1,536	2,048	2,560
Attention heads (dim 128)	12	16	20
Context length	2,048	2,048	2,048
Vocabulary (BPE)	32,768	32,768	32,768
Parameters, total	1.38B	2.82B	4.99B
of which value embeddings	604M	1.07B	1.68B
Scaling parameters (matrices + head)	730M	1.68B	3.23B
Training tokens	5.84B	13.4B	25.8B
Batch size (tokens)	1,048,576	2,097,152	2,097,152
Optimizer steps	5,568	6,400	12,320
Training compute (FLOPs)	2.8 × 10¹⁹	1.4 × 10²⁰	5.3 × 10²⁰
Compute relative to Brain 1	1×	5.2×	19×
Pretraining time, one 8×H100 node	~2 h	~10 h (est.)	~38 h (est.)
Pretraining cost at $27.92–31.92/h	$56–64	$290–330 (est.)	$1,060–1,210 (est.)

Training recipe

Optimizer: Muon for weight matrices, AdamW for embeddings and scalars
Learning rate: 40-step warmup, constant, then linear decay over the final 65% of steps to 5% of peak
Precision: FP8 matrix multiplies on H100, bf16 elsewhere
Attention: Sliding window pattern SSSL (three quarter-context layers, then one full-context layer)
Pretraining data: NVIDIA ClimbMix web text
Fine-tuning data: SmolTalk conversations, MMLU auxiliary train, GSM8K, plus 208 identity conversations × 8 epochs
RL: GSM8K, reward for a correct final answer

Growth-spurt protocol

Schedule: A pretraining spurt every 6 hours, 12 in total, over days 1–3
Resumption: --resume-from-step restores weights, optimizer state and data-loader position
Stopping: --stop-at-step ends a spurt without changing the full-run learning-rate schedule
After each spurt: Validation pass, grading pass, checkpoint kept, report posted
Hardware: One rented 8×H100 SXM node per spurt, released afterwards
Reproducibility: Fixed seeds and data order. Every checkpoint is kept.

Trading curriculum

Objective

The goal is a model that trades profitably. It gets there the way a person would: language first, then arithmetic, then reading markets, then practice without money, then a small budget of its own. Each step starts only when the previous one clears its bar. Every result, good or bad, is published here.

**Table 7.** Curriculum toward trading.
#	Course	How it's taught	Passes when	Status
1	Language	Pretraining on web text	First sentence and first fact milestones	Not started
2	Conversation	Supervised fine-tuning on chats, multiple choice and word problems	Knows its name; answers in turns	Not started
3	Arithmetic	Reinforcement learning on GSM8K	Math score above a threshold set after the first report card	Not started
4	Reading markets	Daily prompts with price moves of Robinhood Chain tokens and recent trader theses from FOMO	Answers parse into a ticker and a side	Planned
5	Paper trading	One pick a day, scored against the market with no money at stake	Four weeks of picks, compared with buy-and-hold NVDA	Planned
6	Trading class	Reinforcement learning on historical market episodes, rewarded by the forward return of each pick	Beats buy-and-hold NVDA on held-out weeks	Planned
7	Live trading	A small weekly budget in its own wallet; one trade a day	Profitable over a quarter, net of costs	Planned

**Table 8.** Rules once it trades.
Decision	One ticker and buy, sell or hold, parsed from the model's own words. Unparseable answers count as hold
Universe	Tokens with at least $50k of pool liquidity. Never $KID
Size and pace	One trade a day, at most 25% of its trading wallet per trade
Execution	Allowlisted router and tokens, with a kill switch
Benchmark	Buy-and-hold NVDA over the same period

**Table 9.** Trading scoreboard.
Period	Picks	Hit rate	Return	vs NVDA
No picks yet. Paper trading starts after course 3.

Compute log

Every spurt runs on one rented 8×H100 SXM node, which is released when the spurt ends.

**Table 10.** Spurts.
Started	Phase	Brain	Target	Cost	Status
No spurts yet.

The $KID token

**Table 11.** Token details.
Name	The Kid
Ticker	$KID
Chain	Robinhood Chain
Paired asset	NVDA (tokenized stock)
Launch venue	Pons
Contract	Not deployed yet

How to buy

Set up a wallet on Robinhood Chain and fund it with ETH on Robinhood Chain.
Open $KID on Pons with the button below, or paste the contract address.
Swap for $KID, then follow its training here.

Contract address Not deployed yet

Buy on Pons Chart X

Verify the address before you buy. There is only one $KID.

Disclosures

Unfiltered outputs. The model is small and trained from scratch. Its outputs are not moderated and may be wrong, strange or offensive.
Estimates. Brain 2 and Brain 3 times and costs are estimates scaled from Brain 1. Rented GPU prices change.
An objective, not a promise. Profitable trading is the research goal. The model may never get there, and nothing on this page promises returns to anyone.
Third-party data. Trader theses used in §8 come from the FOMO API and are written by independent traders.
Not financial advice. $KID is a meme coin with no intrinsic value or expectation of financial return. It exists for entertainment. Do your own research.

References

Karpathy, A. nanochat. MIT license. Training harness, sizing code and GPT-2 CORE reference.
ClimbMix 400B (shuffled). Pretraining corpus.
smol-smoltalk. Conversation data for fine-tuning.
MMLU and GSM8K. Fine-tuning, RL and evaluation.
Power Lines (arXiv 2505.13738). Batch-size scaling used by nanochat's sizing code.
RunPod and Lambda GPU pricing, checked Sep 24, 2026.
FOMO API. Trade theses for §9.