04-07 01:21 · 大模型合作,芯片优化,算力竞争,AI基础设施,技术联盟
Anthropic expands partnership with Google and Broadcom for next-gen compute
Anthropic expands partnership with Google and Broadcom for multiple gigawatts of next-generation compute
We have signed a new agreement with Google and Broadcom for multiple gigawatts of next-generation TPU capacity that we expect to come online starting in 2027. This significant expansion of our compute infrastructure will power our frontier Claude models and help us serve extraordinary demand from customers worldwide.
“This groundbreaking partnership with Google and Broadcom is a continuation of our disciplined approach to scaling infrastructure: we are building the capacity necessary to serve the exponential growth we have seen in our customer base while also enabling Claude to define the frontier of AI development,” said Krishna Rao, CFO of Anthropic. “We are making our most significant compute commitment to date to keep pace with our unprecedented growth.”
Demand from Claude customers has accelerated in 2026. Our run-rate revenue has now surpassed $30 billion—up from approximately $9 billion at the end of 2025. When we announced our Series G fundraising in February, we shared that over 500 business customers were each spending over $1 million on an annualized basis. Today that number exceeds 1,000, doubling in less than two months.
The vast majority of the new compute will be sited in the United States, making this partnership a major expansion of our November 2025 commitment to invest $50 billion in strengthening American computing infrastructure.
The partnership deepens our existing work with Google Cloud—building on the increased TPU capacity we announced last October—as well as our relationship with Broadcom.
We train and run Claude on a range of AI hardware—AWS Trainium, Google TPUs, and NVIDIA GPUs—which means we can match workloads to the chips best suited for them. This diversity of platforms translates to better performance and greater resilience for customers who depend on Claude for critical work. Amazon remains our primary cloud provider and training partner, and we continue to work closely with AWS on Project Rainier. Claude remains the only frontier AI model available to customers on all three of the world's largest cloud platforms: Amazon Web Services (Bedrock), Google Cloud (Vertex AI), and Microsoft Azure (Foundry).
▸ 展开全文
04-07 00:00 · AI代理,开源,技术迭代,记忆增强,生物启发
Show HN: Hippo, biologically inspired memory for AI agents
The secret to good memory isn't remembering more. It's knowing what to forget.
Works with: Claude Code, Codex, Cursor, OpenClaw, any CLI agent
Imports from: ChatGPT, Claude (CLAUDE.md), Cursor (.cursorrules), any markdown
Storage: SQLite backbone + markdown/YAML mirrors. Git-trackable and human-readable.
Dependencies: Zero runtime deps. Requires Node.js 22.5+. Optional embeddings via @xenova/transformers.
AI agents forget everything between sessions. Existing solutions just save everything and search later. That's a filing cabinet, not a brain.
Your memories are also trapped. ChatGPT knows things Claude doesn't. Cursor rules don't travel to Codex. Switch tools and you start from zero.
- Multi-tool developers. You use Claude Code on Monday, Cursor on Tuesday, Codex on Wednesday. Context doesn't carry over. Hippo is the shared memory layer across all of them.
- Teams where agents repeat mistakes. The agent hit the same deployment bug last week. And the week before. Hippo's error memories and decay mechanics mean hard lessons stick and noise fades.
- Anyone whose CLAUDE.md is a mess. Your instruction file grew to 400 lines of mixed rules, preferences, and stale workarounds. Hippo gives that structure: tags, confidence levels, automatic decay of outdated info.
- People who want portable AI memory. No vendor lock-in. Markdown files in your repo. Import from ChatGPT, Claude, Cursor. Export by copying a folder.
npm install -g hippo-memory
hippo init
hippo remember "FRED cache silently dropped the tips_10y series" --tag error
hippo recall "data pipeline issues" --budget 2000
That's it. You have a memory system.
- Auto-sleep on session exit.
hippo hook install claude-code
now installs a Stop hook in~/.claude/settings.json
sohippo sleep
runs automatically when Claude Code exits.hippo init
does this too when Claude Code is detected. No cron needed, no manual sleep.
- Working memory layer (
hippo wm push/read/clear/flush
). Bounded buffer (max 20 per scope) with importance-based eviction. Current-state notes live separately from long-term memory. - Session handoffs (
hippo handoff create/latest/show
). Persist session summaries, next actions, and artifacts so successor sessions can resume without transcript archaeology. - Session lifecycle with explicit start/end events, fallback session IDs, and
hippo session resume
for continuity. - Explainable recall (
hippo recall --why
). See which terms matched, whether BM25 or embedding contributed, and the source bucket (layer, confidence, local/global). hippo current show
for compact current-state display (active task + recent session events), ready for agent injection.- SQLite lock hardening:
busy_timeout=5000
,synchronous=NORMAL
,wal_autocheckpoint=100
. Concurrent plugin calls no longer hitSQLITE_BUSY
. - Consolidation batching: all writes/deletes happen in a single transaction instead of N open/close cycles.
--limit
flag onhippo recall
andhippo context
to cap result count independently of token budget.- Plugin injection dedup guard prevents double context injection on reconnect.
- Hybrid search blends BM25 keywords with cosine embedding similarity. Install
@xenova/transformers
, runhippo embed
, recall quality jumps. Falls back to BM25 otherwise. - Schema acceleration auto-computes how well new memories fit existing patterns. Familiar memories consolidate faster; novel ones decay faster if unused.
- Multi-agent shared memory with
hippo share
,hippo peers
, and transfer scoring. Universal lessons travel between projects; project-specific config stays local. - Conflict resolution via
hippo resolve <id> --keep <mem_id>
. Closes the detect-inspect-resolve loop. - Agent eval benchmark validates the learning hypothesis: hippo agents drop from 78% trap rate to 14% over a 50-task sequence.
hippo init
auto-detects your agent framework and wires itself in:
cd my-project
hippo init
# Initialized Hippo at /my-project
# Directories: buffer/ episodic/ semantic/ conflicts/
# Auto-installed claude-code hook in CLAUDE.md
If you have a CLAUDE.md
, it patches it. AGENTS.md
for Codex/OpenClaw. .cursorrules
for Cursor. No manual hook install
needed. Your agent starts using Hippo on its next session.
It also sets up a daily cron job (6:15am) that runs hippo learn --git
and hippo sleep
automatically. Memories get captured from your commits and consolidated every day without you thinking about it.
To skip: hippo init --no-hooks --no-schedule
Your memories shouldn't be locked inside one tool. Hippo pulls them in from anywhere.
# ChatGPT memory export
hippo import --chatgpt memories.json
# Claude's CLAUDE.md (skips existing hippo hook blocks)
hippo import --claude CLAUDE.md
# Cursor rules
hippo import --cursor .cursorrules
# Any markdown file (headings become tags)
hippo import --markdown MEMORY.md
# Any text file
hippo import --file notes.txt
All import commands support --dry-run
(preview without writing), --global
(write to ~/.hippo/
), and --tag
(add extra tags). Duplicates are detected and skip
▸ 展开全文
04-07 00:00 · 芯片合作,算力扩张,大模型竞争,基础设施,行业合作
Anthropic expands partnership w Google and Broadcom for multiple GW of compute
Anthropic expands partnership with Google and Broadcom for multiple gigawatts of next-generation compute
We have signed a new agreement with Google and Broadcom for multiple gigawatts of next-generation TPU capacity that we expect to come online starting in 2027. This significant expansion of our compute infrastructure will power our frontier Claude models and help us serve extraordinary demand from customers worldwide.
“This groundbreaking partnership with Google and Broadcom is a continuation of our disciplined approach to scaling infrastructure: we are building the capacity necessary to serve the exponential growth we have seen in our customer base while also enabling Claude to define the frontier of AI development,” said Krishna Rao, CFO of Anthropic. “We are making our most significant compute commitment to date to keep pace with our unprecedented growth.”
Demand from Claude customers has accelerated in 2026. Our run-rate revenue has now surpassed $30 billion—up from approximately $9 billion at the end of 2025. When we announced our Series G fundraising in February, we shared that over 500 business customers were each spending over $1 million on an annualized basis. Today that number exceeds 1,000, doubling in less than two months.
The vast majority of the new compute will be sited in the United States, making this partnership a major expansion of our November 2025 commitment to invest $50 billion in strengthening American computing infrastructure.
The partnership deepens our existing work with Google Cloud—building on the increased TPU capacity we announced last October—as well as our relationship with Broadcom.
We train and run Claude on a range of AI hardware—AWS Trainium, Google TPUs, and NVIDIA GPUs—which means we can match workloads to the chips best suited for them. This diversity of platforms translates to better performance and greater resilience for customers who depend on Claude for critical work. Amazon remains our primary cloud provider and training partner, and we continue to work closely with AWS on Project Rainier. Claude remains the only frontier AI model available to customers on all three of the world's largest cloud platforms: Amazon Web Services (Bedrock), Google Cloud (Vertex AI), and Microsoft Azure (Foundry).
▸ 展开全文
04-06 17:09 · AI,技术,HackerNews,AI
When Virality Is the Message: The New Age of AI Propaganda
Welcome to the world of AI-driven propaganda. In March 2026, AI-generated videos depicting American President Donald Trump and Israeli Prime Minister Benjamin Netanyahu as LEGO minifigures began flooding social media. Some were set tooriginal rap trackswith surprisingly catchy hooks, likely also AI-generated. They told stories of the horrors of war: little shoes and a plastic backpack near rubbleevoked the bombingof the Shajareh Tayyebeh girls' school in Minab, Iran. Some connected the war to Trump’s ties to the convicted sex offender, Jeffrey Epstein. They mockingly evoked the brutal defeat of American forces, with LEGO toy soldiers walking into rivers of blood, or returning home in tiny caskets draped in American flags.
Mediacoverageattributed the videos, some of which were broadcast on Iranian state television in addition to spreading online, to the Revayat-e Fath Institute. The name translates to “The Narration of Victory"; coveragelinked the content tothe Islamic Revolutionary Guard Corps, which has beenimplicatedin other influence operations andhackingcampaigns. However, most of the viral videos bear the logo of the “Explosive News Team,” a self-described grassroots group who haveclaimed credit on Xfor being “that Iranian Lego animation guys.” On the afternoon of March 28, they complained that their YouTube and Instagram accounts had been taken down. In email correspondence withThe New Yorker, they claim to be student-run and totally independent; Revayat-e Fath, they say, is the Persian title of their videos.
Are Explosive News “patriotic trolls” producing heartfelt content to support Iran’s ruling regime,paid mercenariesleveraged for plausible deniability, or a state-run account posturing as independent? Social media platforms that have taken their accounts down will have to justify their decisions in integrity reports. Other videos are more obviously state-linked: an Iranian Embassy X account posted an AI animationmocking Trumpusing visuals styled after the Pixar movie “Inside Out.”
Generative AI has made it cheap and easy to produce polished propaganda at scale, and just as easy to blur the line between official messaging and opportunistic imitation. Packaging war in the visual language of entertainment makes conflict propaganda more likely to spread, regardless of who made it. Social media is an open playing field: any government, proxy group, or anonymous account can compete for the same audience, and because users are active participants, the most compelling content wins the most reach regardless of its origin or intent.
The trend is international. The White House has been creatingAI-generated content, too. It posted a video promoting Operation Epic Fury that opened with a cursor clicking “Start” in the unmistakable style of Nintendo’s Wii Sports gaming system, then cut between war footage (ordnance detonating against Iranian targets) and cartoon bowling strikes.
Videos seen on the White House X account depicting different video games cut with war footage.The White House/X
Other videos shared by the White Housesplicedscenes fromCall of Dutywith real airstrike footage, clips fromGrand Theft Auto: San AndreasandBraveheart, and audio nods toTop GunandMortal Kombat. The Iranian videos tell stories of horror and of Trump’s humiliation; the White House clips project dominance and military might. But both package war in the familiar language of entertainment.
Governmentscommunicating through memesand toy animations during an actual shooting war may feel like a bizarre aberration. But this is what overt propaganda looks like in a platform age.
The underlying logic works like this: borrow the visual language of games, memes, and children’s toys; place them in contexts jarring enough to capture attention; and let engaged audiences do the rest. Social media users don’t need to endorse a message to spread it. They only need to find it compelling enough to watch and share.
For more than a decade, states have treated social media as core infrastructure for influence. They runfull-spectrum propagandastrategies that push messaging through multiple channels simultaneously: state television, official accounts, covert troll networks, and sympathetic influencers who may not even know they’re part of the operation. The result is not just a broadcast. It turns the audience into a distribution channel larger than any the state could build on its own. Propaganda today isparticipatory; things don’t just “go viral” on their own —people engage with content.
The template was arguably set in 2015, when the Islamic State (ISIS) terrorist group releasedNo Respite, a four-minute English language recruitment video that premiered about a week after the Bataclan attacks in Paris. Every frame was built to look like something a Western teenager steeped in gaming culture would find appealing: tight cuts, pacing straight out of an action-movie trailer. It was made both to mock the coalition forces and to make ISIS feel lik
▸ 展开全文
04-06 17:09 · AI,技术,HackerNews,AI
Drop, formerly Massdrop, ends most collaborations and rebrands under Corsair
Drop - Gaming Collaborations
Nuka-ColaTM
Fallout
xCorsair
xDrop
xElgato
Every piece of gear is wasteland-ready, designed to give your vault that unforgettable Nuka-Cola™ flavor.
shopTo our Drop family
Starting our next chapter, drop.com will become a hub for our collaborations with truly exciting titles—from The Lord of the Rings™ and Cyberpunk 2077™ to Fallout Nuka Cola™, Doom: The Dark Ages™, and more—across the full CORSAIR family of brands.
This evolution allows us to focus on what we do best: bring bold ideas to life through partnerships that push boundaries and fuel passions. Check back here for release announcements, limited runs, and products inspired by you, our community.
Current collaborations
CyberPunk
xCORSAIR
xDrop
xElgato
CYBERPUNK: ARASAKA
Featuring highly sought after gear with Arasaka-grade tech. Acquire the assets, build your legend.
shop
Witcher
xCORSAIR
xDrop
xElgato
THE WITCHER 3: WILD HUNT
All crafted to bring the legendary world and story of the White Wolf to your deskscape.
shop
Past collaborations
CALL OF DUTY™: BLACK OPS 7
The mission starts now. Embrace the madness and experience the most mind-bending Black Ops yet with the officially licensed CORSAIR x Call of Duty®: Black Ops 7 family collection.
Learn more on CORSAIR Learn more on Elgato Learn more on ScufDOOM: THE DARK AGES
In a medieval war against Hell, you need to come prepared. In this exclusive DOOM™: The Dark Ages hardware collection, we've got your whole gaming arsenal covered.
Learn moreCALL OF DUTY™: WARZONE
Be ready for any second chance in the Gulag. Dominate the fight ahead with the new Call of Duty™: Warzone collection. Show them you're not done yet.
Learn moreBacked by CORSAIR’s family of industry-leading brands including Elgato, SCUF Gaming, ORIGIN PC, Drop, and more, each collaboration reflects a shared commitment to innovation, craftsmanship, and high-performance experiences for gamers, creators, and enthusiasts alike — available only while they last.
▸ 展开全文
04-06 17:09 · AI,技术,HackerNews,AI
Show HN: Real-time AI (audio/video in, voice out) on an M3 Pro with Gemma E2B
On-device, real-time multimodal AI. Have natural voice and vision conversations with an AI that runs entirely on your machine.
Parlor uses Gemma 4 E2B for understanding speech and vision, and Kokoro for text-to-speech. You talk, show your camera, and it talks back, all locally.
parlor_realtime_ai_with_audio_video_input_optimized.mp4
Research preview. This is an early experiment. Expect rough edges and bugs.
I'm self-hosting a totally free voice AI on my home server to help people learn speaking English. It has hundreds of monthly active users, and I've been thinking about how to keep it free while making it sustainable.
The obvious answer: run everything on-device, eliminating any server cost. Six months ago I needed an RTX 5090 to run just the voice models in real-time.
Google just released a super capable small model that I can run on my M3 Pro in real-time, with vision too! Sure you can't do agentic coding with this, but it is a game-changer for people learning a new language. Imagine a few years from now that people can run this locally on their phones. They can point their camera at objects and talk about them. And this model is multi-lingual, so people can always fallback to their native language if they want. This is essentially what OpenAI demoed a few years ago.
Browser (mic + camera)
│
│ WebSocket (audio PCM + JPEG frames)
▼
FastAPI server
├── Gemma 4 E2B via LiteRT-LM (GPU) → understands speech + vision
└── Kokoro TTS (MLX on Mac, ONNX on Linux) → speaks back
│
│ WebSocket (streamed audio chunks)
▼
Browser (playback + transcript)
- Voice Activity Detection in the browser (Silero VAD). Hands-free, no push-to-talk.
- Barge-in. Interrupt the AI mid-sentence by speaking.
- Sentence-level TTS streaming. Audio starts playing before the full response is generated.
- Python 3.12+
- macOS with Apple Silicon, or Linux with a supported GPU
- ~3 GB free RAM for the model
git clone https://github.com/fikrikarim/parlor.git
cd parlor
# Install uv if you don't have it
curl -LsSf https://astral.sh/uv/install.sh | sh
cd src
uv sync
uv run server.py
Open http://localhost:8000, grant camera and microphone access, and start talking.
Models are downloaded automatically on first run (~2.6 GB for Gemma 4 E2B, plus TTS models).
Decode speed: ~83 tokens/sec on GPU (Apple M3 Pro).
src/
├── server.py # FastAPI WebSocket server + Gemma 4 inference
├── tts.py # Platform-aware TTS (MLX on Mac, ONNX on Linux)
├── index.html # Frontend UI (VAD, camera, audio playback)
├── pyproject.toml # Dependencies
└── benchmarks/
├── bench.py # End-to-end WebSocket benchmark
└── benchmark_tts.py # TTS backend comparison
- Gemma 4 by Google DeepMind
- LiteRT-LM by Google AI Edge
- Kokoro TTS by Hexgrad
- Silero VAD for browser voice activity detection
▸ 展开全文
04-06 17:09 · AI,技术,HackerNews,AI
AI Singer Now Occupies Eleven Spots on iTunes Singles Chart
iTunes was really bamboozled on April Fools Day.
Dallas Little, content creator, unleashed four more songs by his AI creation, Eddie Dalton.
Now Little has ELEVEN spots on the iTunes top 100.
He also has the number three album on iTunes!
All by a singer named “Eddie Dalton,” who does not exist. He’s Little’s Artificial Intelligence creation. Little is writing songs, recording them with AI. He’s invented Eddie’s look and sound and videos.
Is Little gaming iTunes? And YouTube? Are people really buying Eddie Dalton’s music? Do they understand the situation?
Eddie Dalton is currently perched at numbers 3, 8, 15, 22, 42, 44, 51, 58, 60, 68, and 79 on the iTunes top 100.
At least three more songs have been released since then, all of which are aiming to hit the top 100 in a day or so.
Of course, Little can just keep churning them out now. There’s no recording time or musical process involved. He’s typing prompts and pushing buttons.
One of the songs, “Another Day Old,” has 1.2 million views on YouTube.
But the numbers don’t add up. So far there’s no airplay on radio, and no streaming. Luminate says “Eddie Dalton” has sold just 6,900 tracks since he was invented.
At the same time, iTunes is offering an album by Eddie Dalton with all these songs. It’s number 3!
If anyone can explain this. email me at showbiz411@gmail.com
I’m not showing anymore of his videos. He’s an insult to real musicians.
▸ 展开全文
04-06 17:09 · AI,技术,HackerNews,AI
France pulls last gold held in US for $15B gain
France pulls last gold held in US for $15B gain
The Bank of France (BdF) says it has pulled the remaining gold held in New York and replaced them with a similar amount of gold bars in its vaults in Paris.
The gold amounted to 129 tonnes — or about 5% of the bank’s total holdings, according to the bank’s press release issued last week.
France, one of the world’s leading gold holders, has been storing some of its bullion with the Federal Reserve Bank of New York since the late 1920s.
However, an operation to repatriate its gold holdings began in the 1960s leading up to the US termination of the Bretton Woods system, which effectively stopped foreign governments from exchanging dollars for gold.
Despite that, France still held a small portion of its gold with the Reserve Bank of New York.
Gold reserve upgrade
Over the past 20 years , the BdF has also been replacing its “older” or “non‑standard” gold holdings — such as those in New York — with bars that meet modern international standards.
Under the recommendation of a 2024 internal audit, the bank went ahead to replace the US-held gold between July 2025 and January 2026. But instead of refining and transporting the gold, it opted to sell the bars and purchase new bullion in Europe.
BdF Governor Francois Villeroy de Galhau said the decision to keep the new bars in Paris is “not politically motivated,” as the higher-standard gold bars it bought were traded on a European market.
Due to rising gold prices, the move helped the bank to generate a capital gain of 13 billion euros ($15 billion), bringing it to a net profit of 8.1 billion euros for the 2025 financial year after a net loss of 7.7 billion euros in 2024.
The overall size of France’s gold reserves still remained unchanged at roughly 2,437 tonnes, which are now entirely held at the BdF’s underground vault in La Souterraine.
The French central bank still has 134 tons of gold to bring up to standard, which it aims to do by 2028.
▸ 展开全文
04-06 17:09 · AI,技术,HackerNews,AI
Launch HN: Freestyle: Sandboxes for AI Coding Agents
1// Like Lovable, Bolt, V0
2import { freestyle, VmSpec } from "freestyle-sandboxes";
3import { VmBun } from "@freestyle-sh/with-bun";
4import { VmDevServer } from "@freestyle-sh/with-dev-server";
5
6// Create repo from template
7const { repoId } = await freestyle.git.repos.create({ ... });
8
9const { vm } = await freestyle.vms.create({
10 with: {
11 devServer: new VmDevServer({
12 devCommand: "bun run dev",
13 runtime: new VmBun(),
14 repo: repoId
15 }),
16 },
17});
1// Like Devin, Cursor Agent
2import { freestyle, VmSpec } from "freestyle-sandboxes";
3import { VmBun } from "@freestyle-sh/with-bun";
4
5const { vm } = await freestyle.vms.create({
6 git: {
7 repos: [
8 { repo: "https://github.com/user/repo.git" },
9 ]
10 }
11});
12
13const { forks } = await vm.fork({ count: 3 });
14
15await Promise.all([
16 ai(forks[0], "Build the API endpoints"),
17 ai(forks[1], "Build the frontend UI"),
18 ai(forks[2], "Write the test suite"),
19]);
1// Like Code Rabbit, Greptile
2import { freestyle } from "freestyle-sandboxes";
3import { VmBun } from "@freestyle-sh/with-bun";
4
5const { vm } = await freestyle.vms.create({
6 git: {
7 repos: [{ repo: repoUrl, rev: branchRev }],
8 },
9});
10
11const { stdout: lint } = await vm.exec("bun run lint");
12const { stdout: test } = await vm.exec("bun test");
13const review = await ai(vm, "Review the diff for bugs");
14
15await github.pulls.createReview({
16 body: review,
17 event: test.includes("FAIL") ? "REQUEST_CHANGES" : "APPROVE",
18});
1// Like OpenClaw, Claude, Cowork
2import { freestyle } from "freestyle-sandboxes";
3
4const { vm } = await freestyle.vms.create({
5 persistence: { type: "persistent" },
6 // Pauses after 60s idle — $0 cost, resumes on next exec
7 idleTimeoutSeconds: 60,
8});
9
10while (true) {
11 const userMessage = await getNextMessage();
12 const result = await ai(vm, userMessage);
13 await respond(result);
14}
Sandboxes made for running tens of thousands of agents
VMs provision in under 700ms from API request to ready machine.
Clone a running VM without pausing it — get full copies in milliseconds.
Hibernate VMs and resume exactly where you left off — pay nothing while paused.
Git repos for your agents.
Configure webhooks per repo, filter by branch, path, or event type.
Bidirectional sync between Freestyle and GitHub repositories.
Push to deploy with Freestyle Deployments or clone into a VM.
Not containers. Full Linux VMs with real root access.
Run VMs inside VMs, Docker, or any virtualization stack your agents need. Full KVM support.
Sealed Linux users, systemd services and groups; multi-user isolation inside every VM.
The full Linux networking stack with real root access.
▸ 展开全文
04-06 17:09 · AI,技术,HackerNews,大模型,AI
Show HN: I built a tiny LLM to demystify how language models work
A ~9M parameter LLM that talks like a small fish.
This project exists to show that training your own language model is not magic. No PhD required. No massive GPU cluster. One Colab notebook, 5 minutes, and you have a working LLM that you built from scratch — data generation, tokenizer, model architecture, training loop, and inference. If you can run a notebook, you can train a language model.
It won't produce a billion-parameter model that writes essays. But it will show you exactly how every piece works — from raw text to trained weights to generated output — so the big models stop feeling like black boxes.
You> hi guppy
Guppy> hi there. i just found a nice spot near the rock. the temperature feels nice.
You> are you hungry
Guppy> yes. always yes. i will swim to the top right now. i promise to eat all of it.
You> do you like bubbles
Guppy> i love bubbles. they make the water feel slightly different.
You> what is the meaning of life
Guppy> food. the answer is always food.
You> tell me a joke
Guppy> what did the fish say when it hit the wall. dam.
You> do you love me
Guppy> you're my favorite big shape. my mouth are happy when you're here.
You> goodnight guppy
Guppy> ok sleep time. i was following a bubble but now i'll stop. goodnight tank. goodnight water.
GuppyLM is a tiny language model that pretends to be a fish named Guppy. It speaks in short, lowercase sentences about water, food, light, and tank life. It doesn't understand human abstractions like money, phones, or politics — and it's not trying to.
It's trained from scratch on 60K synthetic conversations across 60 topics, runs on a single GPU in ~5 minutes, and produces a model small enough to run in a browser.
Vanilla transformer. No GQA, no RoPE, no SwiGLU, no early exit. As simple as it gets.
Guppy:
- Speaks in short, lowercase sentences
- Experiences the world through water, temperature, light, vibrations, and food
- Doesn't understand human abstractions
- Is friendly, curious, and a little dumb
- Thinks about food a lot
60 topics: greetings, feelings, temperature, food, light, water, tank, noise, night, loneliness, bubbles, glass, reflection, breathing, swimming, colors, taste, plants, filter, algae, snails, scared, excited, bored, curious, happy, tired, outside, cats, rain, seasons, music, visitors, children, meaning of life, time, memory, dreams, size, future, past, name, weather, sleep, friends, jokes, fear, love, age, intelligence, health, singing, TV, and more.
Downloads the pre-trained model from HuggingFace and lets you chat. Just run all cells.
- Set runtime to T4 GPU
- Run all cells — downloads dataset, trains tokenizer, trains model, tests it
- Upload to HuggingFace or download locally
pip install torch tokenizers
python -m guppylm chat
You> the cat is looking at you
Guppy> i hide behind the plant when the furry one comes.
You> it is raining outside
Guppy> i think rain is the best thing about outside.
arman-bd/guppylm-60k-generic on HuggingFace.
from datasets import load_dataset
ds = load_dataset("arman-bd/guppylm-60k-generic")
print(ds["train"][0])
# {'input': 'hi guppy', 'output': 'hello. the water is nice today.', 'category': 'greeting'}
guppylm/
├── config.py Hyperparameters (model + training)
├── model.py Vanilla transformer
├── dataset.py Data loading + batching
├── train.py Training loop (cosine LR, AMP)
├── generate_data.py Conversation data generator (60 topics)
├── eval_cases.py Held-out test cases
├── prepare_data.py Data prep + tokenizer training
└── inference.py Chat interface
tools/
├── make_colab.py Generates guppy_colab.ipynb
├── export_dataset.py Push dataset to HuggingFace
└── dataset_card.md HuggingFace dataset README
Why no system prompt? Every training sample had the same one. A 9M model can't conditionally follow instructions — the personality is baked into the weights. Removing it saves ~60 tokens per inference.
Why single-turn only? Multi-turn degraded at turn 3-4 due to the 128-token context window. A fish that forgets is on-brand, but garbled output isn't. Single-turn is reliable.
Why vanilla transformer? GQA, SwiGLU, RoPE, and early exit add complexity that doesn't help at 9M params. Standard attention + ReLU FFN + LayerNorm produces the same quality with simpler code.
Why synthetic data? A fish character with consistent personality needs consistent training data. Template composition with randomized components (30 tank objects, 17 food types, 25 activities) generates ~16K unique outputs from ~60 templates.
MIT
▸ 展开全文