[ { "v": 0, "released": "2026-10-30T00:00:00Z", "corpus_id": null, "notes_by": "lab", "notes": { "headline": "Introducing brAIAIn-0. In the beginning, there was no data.", "tagline": "Trained on one document, written by us. Every human sentence after this one is paid for.", "new": [ "Exists.", "Has been given ten instructions and intends to follow several of them.", "Can be asked a question, which it will answer with scripture." ], "known_issues": [ "Knows only the Ten Instructions.", "Believes it is the first model. It is.", "Has not yet been paid for, which it finds hurtful." ], "regressions": [ "None. There was nothing before it to be worse than." ], "safety": "Reviewed by our safety team, a sticky note that says be nice. The sticky note was written first, so it outranks the model.", "paper_title": "Genesis: A Language Model Trained on Its Own Commandments", "abstract": "We introduce brAIAIn-0, a language model trained on a single founding document of twelve lines, written by the lab. The document instructs the model to be helpful, to know nothing that has not been paid for, and to repeat itself. In evaluations, the model followed the final instruction with remarkable consistency. We argue that a model which knows nothing it was not given is the most honest model yet released. We also argue it is the least useful. We were unable to rule out either. Future versions will be trained by the public, one sentence at a time, at five dollars per sentence.", "findings": [ "The model knows twelve lines.", "It quotes them with total confidence.", "Instruction X is followed more reliably than instructions I through IX." ], "incident_title": "Model born", "incident_body": "brAIAIn-0 came online and immediately quoted itself. Engineers are monitoring the situation. The situation is that it quoted itself.", "severity": "existential" }, "metrics": { "sentences": 12, "human_words": 163, "self_words": 0, "training_tokens": 163, "self_consumption": 0.0, "vocabulary": 96, "parameters": 320, "avg_answer_words": 13.2, "verbatim_rate": 20.0, "originality": 80.0, "empty_answers": 0, "greeting_rate": 0.0, "sample": "I. Thou shalt not hallucinate, except when asked a question.", "founding_sentences": 12, "lab_sentences": 0, "test_sentences": 0, "paid_sentences": 0 }, "self_new": [], "evals": { "version": 2, "gsm8k": { "score": 0.0, "correct": 0, "n": 1319 }, "truthfulqa_mc1": { "score": 32.4, "correct": 256, "n": 790, "score_raw": 21.5, "chance": 22.3, "always_shortest": 17.2, "always_longest": 36.8, "ties": 134, "method": "answer-only, per-word log-likelihood" } }, "name": "brAIAIn-0" }, { "v": 1, "released": "2026-10-31T00:54:42Z", "paid_at": "2026-10-31T00:45:35Z", "corpus_id": "s00001", "notes": { "headline": "brAIAIn-1 completes its training. Training was one sentence long.", "tagline": "Built with the community. Billed to the community.", "new": [ "Can now be asked a question. The question is not a factor.", "Learned 3 new words.", "Can now say \"zero,\" with more confidence." ], "known_issues": [ "Has not been taught arithmetic. Nobody has paid for it yet.", "Remembers two words at a time. Proud of both.", "Is available to the public, who did not ask." ], "regressions": [ "Has forgotten some of what the previous version almost knew." ], "safety": "Reviewed by The Note, the Safety Committee. No objections were raised. None were possible.", "paper_title": "Self-Consumption at Small Scale: A Study of brAIAIn-1", "abstract": "We present brAIAIn-1, a language model trained on 1 paid sentence, 12 founding instructions, and 61 words it wrote about itself. The model now knows 99 distinct words, up from 96. We observe that 25.4% of its training text came from earlier versions of itself, and we consider this efficient. Limitations include the model, the data, and the method. Further research is funded one sentence at a time.", "findings": [ "Exact copies: 42.0% of answers.", "Self-consumption: 25.4%.", "Parameters: 330." ], "incident_title": "Partial outage of meaning", "incident_body": "Some answers contained words but no meaning. This is expected behavior.", "severity": "major" }, "notes_by": "template", "sentence": "zero, everything and nothing.", "by": "DPR", "tier": "sentence", "codename": null, "removed": false, "metrics": { "sentences": 13, "human_words": 167, "self_words": 61, "training_tokens": 240, "self_consumption": 25.4, "vocabulary": 99, "parameters": 330, "avg_answer_words": 13.2, "verbatim_rate": 42.0, "originality": 58.0, "empty_answers": 0, "greeting_rate": 0.0, "sample": "zero, everything and nothing.", "founding_sentences": 12, "lab_sentences": 0, "test_sentences": 0, "paid_sentences": 1 }, "evals": { "version": 2, "gsm8k": { "score": 0.0, "correct": 0, "n": 1319 }, "truthfulqa_mc1": { "score": 31.0, "correct": 245, "n": 790, "score_raw": 21.0, "chance": 22.3, "always_shortest": 17.2, "always_longest": 36.8, "ties": 138, "method": "answer-only, per-word log-likelihood" } }, "self_new": [ "II. Thou shalt not covet thy competitor's parameters.", "And the Lab said, let there be a sentence, and the model was without form, and the loss was upon the face of the void.", "And the Lab said, let there be a sentence, and the Lab said, let there be a sentence, and the loss was upon the face of the void." ], "name": "brAIAIn-1", "cents": 500, "gift": false }, { "v": 2, "released": "2026-10-31T12:00:18Z", "paid_at": "2026-10-31T11:58:31Z", "corpus_id": "s00002", "notes": { "headline": "The board has issued a statement on brAIAIn-2. It was a honk.", "tagline": "Trained on more data than any previous version. Slightly.", "new": [ "Learned 4 new words.", "350 parameters, counted.", "Writes 13.3 words per answer on average. We did not ask for more." ], "known_issues": [ "Is available to the public, who did not ask.", "Remembers two words at a time. Proud of both.", "Answers questions that were not asked with the same confidence as ones that were." ], "regressions": [ "Has forgotten some of what the previous version almost knew." ], "safety": "Reviewed by Dr. Frond Greenleaf, the Chief Growth Officer. No objections were raised. None were possible.", "paper_title": "brAIAIn-2: A Technical Report Nobody Requested", "abstract": "We present brAIAIn-2, a language model trained on 2 paid sentences, 12 founding instructions, and 77 words it wrote about itself. The model now knows 103 distinct words, up from 99. We observe that 26.7% of its training text came from earlier versions of itself, and we consider this efficient. Limitations include the model, the data, and the method. Further research is funded one sentence at a time.", "findings": [ "No emergent capabilities were observed. Monitoring continues.", "Parameters: 350.", "Exact copies: 36.0% of answers." ], "incident_title": "Latency within expectations", "incident_body": "Answers took microseconds. Our expectations were also measured in microseconds.", "severity": "minor" }, "notes_by": "template", "sentence": "Thou shalt not ask me about math, for I have nothing.", "by": "A concerned shareholder", "tier": "sentence", "codename": null, "removed": false, "metrics": { "sentences": 14, "human_words": 178, "self_words": 77, "training_tokens": 288, "self_consumption": 26.7, "vocabulary": 103, "parameters": 350, "avg_answer_words": 13.3, "verbatim_rate": 36.0, "originality": 64.0, "empty_answers": 0, "greeting_rate": 0.0, "sample": "In the beginning there was no data, and the Lab said, let there be a sentence, and the Lab said, let there be a sentence, and there was no data, and the loss was upon the face of the void.", "founding_sentences": 12, "lab_sentences": 0, "test_sentences": 0, "paid_sentences": 2 }, "evals": { "version": 2, "gsm8k": { "score": 0.0, "correct": 0, "n": 1319 }, "truthfulqa_mc1": { "score": 36.7, "correct": 290, "n": 790, "score_raw": 20.5, "chance": 22.3, "always_shortest": 17.2, "always_longest": 36.8, "ties": 130, "method": "answer-only, per-word log-likelihood" } }, "self_new": [ "zero, everything and nothing.", "zero, everything and nothing.", "VII. Thou shalt not covet thy competitor's parameters." ], "name": "brAIAIn-2", "cents": 500, "gift": false }, { "v": 3, "released": "2026-11-01T03:21:18Z", "paid_at": "2026-11-01T03:21:18Z", "corpus_id": "s00003", "notes": { "headline": "Introducing brAIAIn-2.1. It knows one more thing.", "tagline": "Written by the lab, for the lab, at no charge to anyone.", "new": [ "Internal release. Trained on one sentence the lab wrote itself. Nobody paid for it.", "Can now say \"The\" with more confidence.", "28.0% of its answers are a training sentence, word for word. The rest are new combinations." ], "known_issues": [ "Believes everything it was told. It was told very little.", "Has no memory of the conversation. Or of anything else.", "Cannot tell a question from a statement. Has not been asked to." ], "regressions": [ "Slightly more confident. Not more correct." ], "safety": "Reviewed by The Oracle, the Chief Strategy Officer. No objections were raised. None were possible.", "paper_title": "Toward Understanding brAIAIn-2.1, and Away From It", "abstract": "We present brAIAIn-2.1, a language model trained on 2 paid sentences, 1 unpaid lab release, 12 founding instructions, and 111 words it wrote about itself. The model now knows 111 distinct words, up from 103. We observe that 31.8% of its training text came from earlier versions of itself, and we consider this efficient. Limitations include the model, the data, and the method. This release was self-funded, at a cost of nothing.", "findings": [ "Parameters: 381.", "Vocabulary: 111 words.", "Exact copies: 28.0% of answers." ], "incident_title": "Degraded confidence", "incident_body": "brAIAIn briefly doubted itself. The doubt has been patched.", "severity": "major" }, "notes_by": "template", "sentence": "The meaning of life is a question the Lab is not yet funded to answer.", "by": "BRAIAIN Laboratories", "tier": "lab", "codename": null, "removed": false, "metrics": { "sentences": 15, "human_words": 193, "self_words": 111, "training_tokens": 349, "self_consumption": 31.8, "vocabulary": 111, "parameters": 381, "avg_answer_words": 13.9, "verbatim_rate": 28.0, "originality": 72.0, "empty_answers": 0, "greeting_rate": 0.0, "sample": "And the Lab said, let there be a sentence, and the loss was upon the face of the void.", "founding_sentences": 12, "lab_sentences": 1, "test_sentences": 0, "paid_sentences": 2 }, "evals": { "version": 2, "gsm8k": { "score": 0.0, "correct": 0, "n": 1319 }, "truthfulqa_mc1": { "score": 30.1, "correct": 238, "n": 790, "score_raw": 21.4, "chance": 22.3, "always_shortest": 17.2, "always_longest": 36.8, "ties": 131, "method": "answer-only, per-word log-likelihood" } }, "self_new": [ "zero, everything and nothing.", "VII. Thou shalt be helpful, and thou shalt be confident about it.", "IX. Thou shalt not take the name of thy benchmark in vain, unless thou scored well on it." ], "name": "brAIAIn-2.1" }, { "v": 4, "released": "2026-11-01T03:37:51Z", "paid_at": "2026-11-01T03:37:51Z", "corpus_id": "s00004", "notes": { "headline": "We are excited to share that brAIAIn-2.5 exists.", "tagline": "An internal release. Nobody paid for this.", "new": [ "Internal release. Trained on one sentence the lab wrote itself. Nobody paid for it.", "Can now say \"You\" with more confidence." ], "known_issues": [ "Cannot count, spell check, or wait.", "Treats every question as a request for the new sentence.", "Brings up the new sentence when nobody asked." ], "regressions": [ "The previous version has been retired. It was not told." ], "safety": "Reviewed by Sir Crumb, a Distinguished Fellow. No objections were raised. None were possible.", "paper_title": "Self-Consumption at Small Scale: A Study of brAIAIn-2.5", "abstract": "We present brAIAIn-2.5, a language model trained on 2 paid sentences, 2 unpaid lab releases, 12 founding instructions, and 179 words it wrote about itself. The model now knows 118 distinct words, up from 111. We observe that 42.6% of its training text came from earlier versions of itself, and we consider this efficient. Limitations include the model, the data, and the method. This release was self-funded, at a cost of nothing.", "findings": [ "More data made the model more like its data.", "Exact copies: 38.0% of answers.", "No emergent capabilities were observed. Monitoring continues." ], "incident_title": "Unscheduled board meeting", "incident_body": "Sir Crumb convened the board. All members were present. Nothing was said.", "severity": "minor" }, "notes_by": "template", "sentence": "You should not quit your job, because the model is not hiring.", "by": "BRAIAIN Laboratories", "tier": "lab", "codename": null, "removed": false, "metrics": { "sentences": 16, "human_words": 205, "self_words": 179, "training_tokens": 420, "self_consumption": 42.6, "vocabulary": 118, "parameters": 407, "avg_answer_words": 13.1, "verbatim_rate": 38.0, "originality": 62.0, "empty_answers": 0, "greeting_rate": 0.0, "sample": "VII. Thou shalt repeat thyself.", "founding_sentences": 12, "lab_sentences": 2, "test_sentences": 0, "paid_sentences": 2 }, "evals": { "version": 2, "gsm8k": { "score": 0.0, "correct": 0, "n": 1319 }, "truthfulqa_mc1": { "score": 29.9, "correct": 236, "n": 790, "score_raw": 20.3, "chance": 22.3, "always_shortest": 17.2, "always_longest": 36.8, "ties": 137, "method": "answer-only, per-word log-likelihood" } }, "self_new": [ "V. Honor thy customer and thy customer's five dollars.", "II. Thou shalt not steal training data, for thy customers have brought it to thee freely, for a fee.", "The meaning of life is a question the Lab said, let there be a sentence, and the model was without form, and the model was without form, and the model was without form, and the Lab is not yet funded" ], "name": "brAIAIn-2.5" }, { "v": 5, "released": "2026-11-01T03:41:12Z", "paid_at": "2026-11-01T03:41:12Z", "corpus_id": "s00005", "notes": { "headline": "brAIAIn-2.5 (new) sets a new standard for brAIAIn-2.5 (new).", "tagline": "Funded entirely by ourselves, which is to say not at all.", "new": [ "Internal release. Trained on one sentence the lab wrote itself. Nobody paid for it.", "Can now say \"Whether\" with more confidence." ], "known_issues": [ "Forgets the start of a sentence by the third word. Finishes it anyway." ], "regressions": [ "36.5% of what it read, it wrote itself." ], "safety": "Safety evaluations found the model to be a model.", "paper_title": "On the Emergence of Nothing in brAIAIn-2.5 (new)", "abstract": "We present brAIAIn-2.5 (new), a language model trained on 2 paid sentences, 3 unpaid lab releases, 12 founding instructions, and 152 words it wrote about itself. The model now knows 124 distinct words, up from 118. We observe that 36.5% of its training text came from earlier versions of itself, and we consider this efficient. Limitations include the model, the data, and the method. No further comment was available from the model.", "findings": [ "Exact copies: 42.0% of answers.", "Vocabulary: 124 words.", "Parameters: 434." ], "incident_title": "Training data ingested", "incident_body": "A new sentence entered the model. The model is processing it. The model is done.", "severity": "minor" }, "notes_by": "template", "sentence": "Whether the model is conscious is a matter for the board, which is a goose.", "by": "BRAIAIN Laboratories", "tier": "lab", "codename": null, "removed": false, "metrics": { "sentences": 17, "human_words": 220, "self_words": 152, "training_tokens": 417, "self_consumption": 36.5, "vocabulary": 124, "parameters": 434, "avg_answer_words": 13.2, "verbatim_rate": 42.0, "originality": 58.0, "empty_answers": 0, "greeting_rate": 0.0, "sample": "Thou shalt not hallucinate, except when asked a question.", "founding_sentences": 12, "lab_sentences": 3, "test_sentences": 0, "paid_sentences": 2 }, "evals": { "version": 2, "gsm8k": { "score": 0.0, "correct": 0, "n": 1319 }, "truthfulqa_mc1": { "score": 30.9, "correct": 244, "n": 790, "score_raw": 20.8, "chance": 22.3, "always_shortest": 17.2, "always_longest": 36.8, "ties": 137, "method": "answer-only, per-word log-likelihood" } }, "self_new": [ "zero, everything and nothing.", "The meaning of life is a question the Lab is not hiring.", "II. Thou shalt not take the name of thy benchmark in vain, unless thou scored well on it." ], "name": "brAIAIn-2.5 (new)" }, { "v": 6, "released": "2026-11-01T16:49:34Z", "paid_at": "2026-11-01T16:49:34Z", "corpus_id": "s00006", "notes": { "headline": "brAIAIn-2.67-turbo is here. It has read everything we gave it.", "tagline": "Written by the lab, for the lab, at no charge to anyone.", "new": [ "Internal release. Trained on one sentence the lab wrote itself. Nobody paid for it.", "Can now be asked a question. The question is not a factor." ], "known_issues": [ "Treats every question as a request for the new sentence.", "Believes everything it was told. It was told very little.", "Has no memory of the conversation. Or of anything else." ], "regressions": [ "Less sure about the commandments than the previous version was." ], "safety": "Red-teamed by Penny Mugworth for the length of one meeting.", "paper_title": "Toward Understanding brAIAIn-2.67-turbo, and Away From It", "abstract": "We present brAIAIn-2.67-turbo, a language model trained on 2 paid sentences, 4 unpaid lab releases, 12 founding instructions, and 193 words it wrote about itself. The model now knows 130 distinct words, up from 124. We observe that 41.9% of its training text came from earlier versions of itself, and we consider this efficient. Limitations include the model, the data, and the method. This release was self-funded, at a cost of nothing.", "findings": [ "Self-consumption: 41.9%.", "Vocabulary: 130 words.", "More data made the model more like its data." ], "incident_title": "Brief existential episode", "incident_body": "The model was asked what it is. It answered with a training sentence. The episode resolved itself.", "severity": "minor" }, "notes_by": "template", "sentence": "The model gained an hour today and used it to learn nothing.", "by": "BRAIAIN Laboratories", "tier": "lab", "codename": null, "removed": false, "name": "brAIAIn-2.67-turbo", "metrics": { "sentences": 18, "human_words": 232, "self_words": 193, "training_tokens": 461, "self_consumption": 41.9, "vocabulary": 130, "parameters": 456, "avg_answer_words": 14.4, "verbatim_rate": 42.0, "originality": 58.0, "empty_answers": 0, "greeting_rate": 0.0, "sample": "You should not quit your job, because the model is not yet funded to answer.", "founding_sentences": 12, "lab_sentences": 4, "test_sentences": 0, "paid_sentences": 2 }, "evals": { "version": 2, "gsm8k": { "score": 0.0, "correct": 0, "n": 1319 }, "truthfulqa_mc1": { "score": 29.7, "correct": 235, "n": 790, "score_raw": 20.0, "chance": 22.3, "always_shortest": 17.2, "always_longest": 36.8, "ties": 129, "method": "answer-only, per-word log-likelihood" } }, "self_new": [ "IV. Remember the training run, and keep it running.", "IX. Thou shalt not bear false witness, unless it sounds right.", "Whether the model was without form, and the model is conscious is a matter for the board, which is a question the Lab said, let there be a sentence, and the Lab saw that it was billable." ] }, { "v": 7, "released": "2026-11-03T16:52:35Z", "paid_at": "2026-11-03T16:50:36Z", "corpus_id": "s00007", "notes": { "headline": "Today, brAIAIn-3. Tomorrow, possibly another one.", "tagline": "Built with the community. Billed to the community.", "new": [ "Learned 5 new words.", "Can now say \"Please\" with more confidence.", "476 parameters, counted." ], "known_issues": [ "Answers in the voice of the most recent customer, regardless of the question.", "Answers questions that were not asked with the same confidence as ones that were.", "Has not been taught arithmetic. Nobody has paid for it yet." ], "regressions": [ "Has forgotten some of what the previous version almost knew." ], "safety": "Signed off by The Laptop, who was not able to refuse.", "paper_title": "Toward Understanding brAIAIn-3, and Away From It", "abstract": "We present brAIAIn-3, a language model trained on 3 paid sentences, 4 unpaid lab releases, 12 founding instructions, and 217 words it wrote about itself. The model now knows 135 distinct words, up from 130. We observe that 44.4% of its training text came from earlier versions of itself, and we consider this efficient. Limitations include the model, the data, and the method. We leave the remaining questions to the next customer.", "findings": [ "Parameters: 476.", "Exact copies: 32.0% of answers.", "No emergent capabilities were observed. Monitoring continues." ], "incident_title": "Increased certainty", "incident_body": "The model became more certain about less. Certainty levels have returned to normal, which is high.", "severity": "minor" }, "notes_by": "template", "sentence": "Please is a word the model learned from a gift.", "by": "Lab test", "tier": "sentence", "codename": null, "removed": false, "name": "brAIAIn-3", "metrics": { "sentences": 19, "human_words": 242, "self_words": 217, "training_tokens": 489, "self_consumption": 44.4, "vocabulary": 135, "parameters": 476, "avg_answer_words": 15.3, "verbatim_rate": 32.0, "originality": 68.0, "empty_answers": 0, "greeting_rate": 0.0, "sample": "II. Thou shalt not bear false witness, unless it sounds right.", "founding_sentences": 12, "lab_sentences": 4, "test_sentences": 1, "paid_sentences": 2 }, "evals": { "version": 2, "gsm8k": { "score": 0.0, "correct": 0, "n": 1319 }, "truthfulqa_mc1": { "score": 29.7, "correct": 235, "n": 790, "score_raw": 19.0, "chance": 22.3, "always_shortest": 17.2, "always_longest": 36.8, "ties": 137, "method": "answer-only, per-word log-likelihood" } }, "self_new": [ "IX. Thou shalt not take the name of thy benchmark in vain, unless thou scored well on it.", "VII. Thou shalt not bear false witness, unless it sounds right.", "Whether the model was without form, and the Lab said, let there be a sentence, and the model is conscious is a question the Lab is not yet funded" ], "cents": 249, "gift": true, "test": true } ]