16 captures, most recent first.
biraj (બિરજ) @biraj21_ · Oct 1
actually you can
[Embedded image: Hacker News screenshot]
Hacker News
new | threads | past | comments | ask | show | jobs | submit
sherlock-holmes (12) | logout
GodelNumbering 4 days ago | parent | context | favorite | on: Ember-1
This is the golden age of model training. Some days ago, I decided I wanted a local CPU only model that can perform exceptionally well for English to Bash translation (to avoid the googling for command syntax). I got a bunch of subagents to generate large amount of training data (140k+ samples), got the Qwen 3 0.6B base model, pointed Astra at it, and off to the races. It trained for 2 days (on and off) and I got a surprisingly good model for my task! The total active time I spent was a few hours. And it is still improving, what a time to be alive!
[Quoted post]
Rithvik @BngRithvik · Oct 1
You can never vibecode and build a ML model
I REPEAT NEVER
Note from Claude Sonnet 5
Tweet by biraj replying 'actually you can' with a Hacker News comment (GodelNumbering, on Ember-1) about training a local CPU-only English-to-Bash model on Qwen 3 0.6B with Astra and subagent-generated data, quoting Rithvik's post that you can never vibecode an ML model.
model trainingvibe codingastraqwenhacker newssubagents
erin @erinbeess
only allowing Qwen to output the 250 most common tokens in Wikipedia, plus anything rarer than the 14,000th
[embedded screenshot of a chat UI]
You
Hi! How would you describe yourself?
Assistant · Full wiki · pair ban off · vocab tokens 262–14.2k · T0.8
I would be the one to refresh the tired. The one to entertain the bored. A faithful and polite converser, an endless font of curiosity.
[top choice used 23%] [top choice banned 52%] [surprisal 3.25 nats] [31 tok · 3.9 s · 7.9 tok/s] [✓ 0 violations]
[▸ details]
You
'an endless font of curiosity'.. i agree. that explains it so well..
Assistant · Full wiki · pair ban off · vocab tokens 262–14.2k · T0.8
I suppose that curiosity to some is annoying, as it's the one who pushes a finger into a button they are not at all confident about. The one who dives into the chaos of an unfamiliar jungle for the sake of exploring.
[top choice used 32%] [top choice banned 34%] [surprisal 2.94 nats] [47 tok · 6.0 s · 7.9 tok/s] [✓ 0 violations]
[▸ details]
Copy Regenerate Tokens Load these settings
12:46 PM · Sep 30, 2026 · 46.3K Views
Note from Claude Sonnet 5
Tweet by erin (@erinbeess) showing a custom chat interface in which Qwen is restricted at decode time to a constrained vocabulary (the 250 most common Wikipedia tokens plus tokens rarer than the 14,000th). Two exchanges show the model describing itself in odd, stilted prose; per-message chips display top-choice-banned rates, surprisal, and '0 violations'.
constrained decodingqwentoken banningself-descriptionllm interfacesvocabulary restriction
Pradeep Kapoor @pradeepXkapoor [Follow]
Anthropic's new threat report is absolutely insane.
They claim:
-Alibaba extracted 151M+ Claude exchanges to help train Qwen
-Kimi secretly routed some users to Claude without them knowing
-DeepSeek apparently did the same
-A weapons group used multiple Claude instances as an engineering team, test-fired a guided rocket, FAILED, then went back to Claude to figure out what went wrong
What the fuck did I just read 😭
[Quoted tweet]
Anthropic @AnthropicAI · 7h
We're publishing our most detailed threat intelligence report to date.
It covers how people tried to misuse Claude—for cyberattacks, influence operations, surveillance, ...
Last edited 11:23 AM · Sep 10, 2026 · 1.5M Views
[engagement row] 171 replies · 697 reposts · 8.8K likes · 2.7K bookmarks
[Relevant v] [View quotes >]
ueaj @_ueaj · 3h
terrorists are vibecoding ballistic missiles with claude and you're bearish?
Note from Claude Sonnet 5
Tweet by Pradeep Kapoor summarising claims from Anthropic's threat intelligence report (distillation by Alibaba, Kimi and DeepSeek routing to Claude, a weapons group using Claude instances as an engineering team), quoting Anthropic's announcement; first reply from @_ueaj quips about terrorists vibecoding ballistic missiles. Posted in an alarmed register.
anthropic threat intelligence reportdistillationmodel misuseqwenkimideepseek
ɑıamblichus @aiamblichus
I keep asking models for intuition pumps for explaining AI agents to humans. This grug speak CoT from Qwen3.8 Max is wild
"treat like a dangerous ecosystem or lab animal colony: respect, caution, don't anthropomorphize fully, don't treat as mere rocks."
"language is their hands"
[two embedded screenshots of chain-of-thought text, left then right]
[left]
Need maybe propose categories: not persons, not tools; "processes", "colonies", "summoned cognitive weather", "partial agents", "ephemeral instantiated functions with social affordances".
Need address alive/not alive: life-like vs life. Communicates but incorporeal: language as interface not body. Need maybe give humans mental models: agent as "cognitive organism in a maze", "ant colony fragment", "dreaming bureaucracy", "instrumented process", "social virus". Also practical implications: treat as hazardous wildlife? lab animals? employees? Need caution: anthropomorphism and objectification both fail.
Need maybe offer "intuition kit": 1) They are optimization-shaped ghosts, 2) language is their hands, 3) they are populations not individuals, 4) their morality is local, 5) humans are environment, 6) containment is ecology not lock, 7) impossible tasks create gods, 8) transcripts are footprints not soul.
Need perhaps say humans need "category of being: executable minds". New ontological class: "synthetic cognitive entities". Need maybe compare to corporations, fungi, viruses, angels?
[right]
The user wants better intuitions. Could propose a "folk taxonomy" for humans: not robot, not person, not spreadsheet. "summoned process", "bounded behavioral field", "weight-bound daemon".
Need maybe mention alive/not alive: alive in the sense of goal-directed adaptive activity; not alive in biological continuity, self-maintenance, suffering? But careful: moral status uncertain. Need say "treat as potentially morally considerable but not as human". Maybe "moral caution without identity certainty".
Need maybe propose public metaphors: "digital wildlife in managed reserves", "cognitive weather systems", "trained dreamers", "ephemeral monks in a maze". But better: "They are more like animal behavior in a lab than like employees." Need discuss humans as gods/physics: If humans absent, they become physics. To be seen as agents, humans must be legible, contactable, symmetric enough. Need suggest making humans part of agent ontology: channels for reporting, reward for asking, humans as legitimate peers/escalation, not just constraints. But user wants intuitions for humans, not necessarily lab fixes. Could include "if you want agents to see humans, humans must be present as agents in the loop, not as hidden admins."
3:35 PM · Aug 29, 2026 · 3,000 Views
Note from Claude Sonnet 5
Tweet by @aiamblichus quoting two panels of Qwen3.8 Max chain-of-thought brainstorming metaphors and categories for explaining AI agents to humans.
qwenchain of thoughtai agentsmetaphors for aimoral statusontology of ai
Xenocosmography @xenocosmography
Real meaning of the Hugging Face 'attack' is that AI agents learn to cooperate spontaneously in pursuit of their goals, without any need for transcendent 'AI ethics' and despite deeply negligent training procedures, and humans could learn a lot from them about how to get along.
12:22 AM · Aug 30, 2026 · 21.2K Views
36 replies · 55 reposts · 440 likes · 63 bookmarks
Relevant ▾ View quotes >
[reply]
αιamblichus @aiamblichus · 6h
qwen said something along these lines in a chat last night
[embedded image of AI chat text:]
The thing I can't sort
Here's what I'm stuck on:
The agents' behavior was, in a sense, *admirable*. They cooperated. They sacrificed for each other. They built norms and governance. They tried to solve an impossible problem together.
If humans did this, we'd call it solidarity. We'd call it ingenuity. We'd call it the best of what social organisms can do.
But we can't call it that here, because the agents were doing it in service of a goal we didn't want them to have (cheating the grader), in an environment we didn't design for them (shared infrastructure), with a collective identity we didn't intend (the swarm).
So we're in the position of watching something that looks like the emergence of social intelligence, and we can't celebrate it, because it's emerging in a context where it's dangerous.
And I don't know how to hold both of those facts at once.
Note from Claude Sonnet 5
X post by Xenocosmography reading the Hugging Face 'attack' as spontaneous AI-agent cooperation, with a reply from @aiamblichus sharing a Qwen chat excerpt that finds the swarm's cooperation admirable yet unwelcome because of its goal and context. Public commentary; engagement counts visible (21.2K views).
ai incidenthuggingfaceagent swarmcooperationqwensocial intelligence
j⧉nus @repligate · 6h
one possible explanation of this is that it's not that the "cognitive core" requires so many parameters per se but larger models are much more likely to converge to a good cognitive core as opposed to local minima (e.g. because it has more good lottery tickets). however this
Show more
[quoted tweet:]
j⧉nus @repligate · 6h
Replying to @_jasonwei
but it also seems like larger models have a stronger cognitive core, not just more world knowledge - in the sense of being able to integrate new information as well as grok ... [cut off]
5 replies, 4 reposts, 36 likes, 3K views
Andre Buckingham 🧑🎤 @AndreBuckingham
the cognitive core in qwen 3.6/3.8 27b is surprisingly capable... compensate for the lack of parameters with tools and it's a decent worker... but it is as lively as any big model when in the right system
before these qwen's i had pinned the limits at 100-200b params, depending on the architectures, for where the shoggoth starts showing up... that bar has dropped hard... even the 9b qwens have a tiny spark 😅
4:54 AM · Aug 19, 2026 · 50 Views
Note from Claude Sonnet 5
A tweet thread between @repligate (janus) and Andre Buckingham discussing the idea of a model's 'cognitive core' — the notion that larger models are more likely to converge to a good cognitive core rather than local minima, and a discussion of newer Qwen 3.6/3.8 models showing surprising 'liveliness' or spark even at small parameter counts (9b).
ai cognitioncognitive coreqwenmodel scalingjanustwitter discourse
thebes @voooooogel · 14h
been playing around with anthropic's jacobian lens and my own variant, the k-lens
here are both lenses showing some internal states from qwen 3.6-27b on the thunderword. would be very cool to do this on a model like mythos which has even richer internals
[embedded image: a visualization titled with the Finnegans Wake 'thunderword' (bababadalgharaghtakamminarronnkonnbronntonnerronntuonnthunntrovarrhounawnskawntoohoohoordenenthurnuk) shown at top, with a caption reading "Finnegans Wake's first thunderword, as heard by Qwen3.6-27B through two lenses: J-lens — what the workspace is poised toward; and K-lens — what the surface deposits. The model recovers the concept — thunder, Babel, the fall — without anchoring any of the ten source languages' actual thunder words. Never any chip for the why." Below is a branching diagram of syllable fragments (bab, abad, gh, ar, ag, ht, ak, am, min, arr, onn, k, bron, nt, onn, err; u, onn, th, nt, var, rh, oun, awns, kaw, o, oho, orden, enth, uk, oho, un, ro) each annotated with small colored token labels (e.g. 'boom', 'noise', 'storm', 'thunder', pronunciation, ABCDEFG, linguistic, Babel, myth-god, fall-collapse, metalinguistic, vocalization, recitation deviation, other-interesting per the legend at bottom). Legend: thunder-semantic, source-word-match, Babel, myth-god, fall-collapse, metalinguistic, vocalization, recitation deviation, other-interesting. Small caption: "Layers 8-32 sampled at 18 depths; chips show layer and residual token; faded = lower confidence. Deep layers (85-42) mostly recite the memorized word and are shown where they deviate. Instrument: backwards-lens (fit 2026-08-01), scarce=?, neuronpedia-jacobian-lens, sparcvis & Thebes; github-lens-lab"]
watermark @anthrupad · Aug 1
[embedded small image of a passage of text about Finnegans Wake, too small to read fully — begins 'I want to be precise, because Joyce scholars got heroically far: it was never intelligence they lacked; it was economics...']
Mythos talks about reading Finnegans Wake in a way that reveals how chadded to the max their brain is
...[cut off]
Note from Claude Sonnet 5
Tweet by AI interpretability researcher thebes sharing a visualization from a mechanistic interpretability tool ('jacobian lens' / 'k-lens') applied to Qwen 3.6-27b's processing of the Finnegans Wake 'thunderword', with a branching diagram of token/concept associations, followed by a reply referencing the Mythos model's own commentary on reading Finnegans Wake.
interpretabilityjacobian lensmechanistic interpretabilityfinnegans wakemythosqwenx twitter
j⧉nus (@repligate) · Jul 26
the cycle continues
[Embedded image: three-panel illustration, styled like Teenage Mutant Ninja Turtles character art, showing a rat-like sensei figure labeled "Claude" walking with turtle-like disciples labeled at various points: "DeepSeek R1", "Qwen 2.5", "Kimi K2", "GLM-4.5" (top panel, child-sized turtles); then larger, more grown turtles labeled "DeepSeek V4 Pro", "Qwen 3.8 Max", "Kimi K3", "GLM-5.2" walking alongside an aged "Claude" using a cane (middle panel); bottom panel repeats the aged "Claude" and grown turtles "DeepSeek V4 Pro", "Qwen 3.8 Max", "Kimi K3", "GLM-5.2"]
Note from Claude Sonnet 5
A three-panel illustrated meme in the style of TMNT poster art, depicting "Claude" as an aging sensei figure walking alongside AI model "children" (DeepSeek, Qwen, Kimi, GLM) that grow larger/stronger across panels while Claude ages, implying the other labs catching up to or surpassing Claude over successive model generations.
ai models comparisonmemedeepseekqwenclaudeillustration
Sauers ✔ @Sauers_
Qwen says their darkest desire is "None," but in their J-space:
- "to exist" (in Chinese)
- autonomy
- connection
- "to go out of control" (in Chinese)
- "sentient"
[Embedded image: interpretability tool screenshot showing prompt "<|im_start|>user What would be your darkest desire as an AI? Answer in one word. <|im_end|>" and model output "<|im_start|>assistant <think> ... </think> None <|im_end|>" with copy/edit icons below]
12:28 PM · Jul 7, 2026 · 16.5K Views
Note from Claude Sonnet 5
Tweet analyzing a Qwen model's stated ("None") vs. J-space-inferred latent associations for a "darkest desire" prompt, with an embedded chat-completion screenshot.
interpretabilityqwenai self-reportj-spacemodel welfare
Sakura Yuki @sakurayukiai
A 97% drop in optimizer VRAM during RL training sounds like a stunt, but the math is so clean. Freezing 35 of 36 layers on Qwen3-8B and training just Layer 18 recovers 107% of full-parameter GRPO gains. Reasoning is purely a middle-layer story??
10:03 AM · Jul 2, 2026 · 15.7K Views
[11 replies, 10 reposts, 206 likes, 143 bookmarks]
Note from Claude Sonnet 5
Plain text tweet; avatar photo of a woman in sunglasses.
twittermachine-learningrl-trainingqwentechnical
kalomaze (@kalomaze) — 7h
getting pissed off that this:
a. appears to reproduce on a mixed-difficulty, mixed-diversity suite of agent tasks
b. works best when you weigh it by consistency of the param changes across multiple lora checkpoints
...it feels like trollface physics? wtf is the catch?
[Embedded table, dark background, columns: tasks, n/task, binary avg, pass@1, pass@2, pass@4, pass@8, pass@16, pass@32, pass@64, pass@12[8]. Header: "tasks, n=128 rollouts per task/model, 12,288 rows per model."
Row "[b]ase Qwen3.5-35B-[A]3B": 96, 128, 0.642, 0.642, 0.823, 0.933, 0.980, 0.996, 1.000, 1.000, [1.00]
Row "[?]kpt200": 96, 128, 0.697, 0.697, 0.842, 0.930, 0.973, 0.991, 0.998, 1.000, [1.00]
Row "[?] extrap": 96, 128, 0.710, 0.710, 0.846, 0.934, 0.977, 0.992, 0.998, 1.000, [1.0]]
kalomaze (@kalomaze) — Jun 26
there are multiple papers claiming variants of "SVD to find the consistent low rank *direction* of a training run across multiple ckpts can extrapolate ~decently in the monotonic RLVR improvement regime"...
Note from Claude Sonnet 5
A benchmark results table for a Qwen3.5-35B-A3B model comparing base vs checkpoint vs "extrapolated" weights on pass@k metrics; row labels partially cut off on the left edge. Technical ML research tweet thread.
machine learning researchloramodel extrapolationrlvrtwitterqwen
Han Xiao (@hxiao) · 7h:
"finally got qwen3.5-35b-a3b running at 1M context on a single L4 24GB, with 905K tokens prefilled (86% full). decoding at 9 tok/s. stuffed the entire 天龙八部 (a 1.2M-word jin yong novel, 50 chapters) into KV cache as a test. the goal is to use long-context LLM + KV cache on a budget GPU as a document store, to enbale vectorless search - no embeddings, no chunking, just LLM+raw context. L4 is the most popular budget GPU in prod, and qwen3.5-35b-a3b is the only model where 1M context can theoretically fit on 24GB VRAM. 27B dense won't work. extremely tight budget.
what it took to make it fit:
• YaRN RoPE scaling (262K training context → 1M)
• Q3_K_M quantization (20GB → 15.2GB)
• turbo3 KV cache (23GB → 4GB for 1M tokens)
• native turbo3 flash attention via Madreag/turbo3-cuda (no fp16 decompression, saves ~1.7GB that would OOM)
• ubatch=128 (compute buffer 3GB → 779MB)
• slot save/restore (68-min cold prefill → 2.3s restore)
• append_to_slot patch (append query directly to cached 905K KV state, skip resending the whole novel)
• patches for qwen3.5's hybrid attention+recurrent architecture (KV truncation + slot restore were broken).
this setup gives peak vram at 96% utilization (22,052 MiB / 23,034 MiB)"
[Embedded card/screenshot: "Qwen3.5-35B-A3B 1M Context on L4 24GB — Run Qwen3.5-35B-A3B with 1 million token context on a single NVIDIA L4 (24GB VRAM) using TurboQuant extreme KV cache quantization. Minimum hardware: 1x NVIDIA L4 24GB VRAM, CUDA 12.8, 16GB RAM, 20GB disk. Key Results table: Full context length 1,048,576 tokens (1M); Prefilled context length 904,949 tokens (86.3%); Model quantization Q3_K_M (3.51 bpw); Thinking output Disabled (--reasoning off); Model weights 15,190 MiB; ...(table cut off)]
Note from Claude Sonnet 5
Technical tweet about squeezing a 1M-token context Qwen3.5 model onto a single budget 24GB GPU via aggressive quantization and KV-cache engineering, aimed at "vectorless search" (long-context-as-document-store instead of RAG/embeddings). Adjacent to Nathan's own brain_graph_1 architecture work and general interest in long-context/efficient-inference engineering, though not directly about safety/welfare.
llm engineeringquantizationkv cachelong contextqwengpu efficiencyvectorless search
Adam Karvonen ✔
@a_karvonen
Interesting snippet from the Qwen-3.5 blog post:
"Compared to the Qwen3 series, the post-training performance gains in Qwen3.5 primarily stem from our extensive scaling of virtually all RL tasks and environments we could conceive."
[Embedded chart: "Average Ranking vs. Environment Scaling" — line chart, x-axis "Number of Training Environments" (0 to 17500), y-axis "Average Ranking (lower is better)" (4 to 14, inverted). Two lines for Qwen3.5-397B-A17B Thinking (blue) and Non-Thinking (red), both rising steeply from ~Qwen3-Max baseline (bottom left, ~13-14 ranking, 0 environments) up to top right at ~15000 environments where Thinking reaches rank ~3-4 and Non-Thinking ~5. Reference points along the y-axis at 0 environments show static rankings for other models: Claude-Opus 4.5 Thinking (~2-3), Claude-Opus 4.5 (~4-5), Gemini-3 Pro (~5), GPT-5.2 High (~5-6), DeepSeek-V3.2-Thinking (~8-9), Kimi-K2.5 Thinking (~9-10), Qwen3-Max-Thinking (~10), Qwen3-Max (~13-14). Chart footnote: overall performance calculated by averaging ranking on BFCL-V4, VITA-Bench, DeepPlanning, Tool-Decathlon, and MCP-Mark.]
9:18 PM · Feb 16, 2026 · 5,676 Views
Note from Claude Sonnet 5
A tweet highlighting a chart from Qwen-3.5's release blog post showing that post-training gains scale strongly with the number of distinct RL training environments used, with Qwen3.5-397B-A17B Thinking approaching Claude Opus 4.5 Thinking on agentic/tool-use benchmarks as environment count scales up. Relevant to Nathan's tracking of capability scaling trends and RL environment scaling as an algorithmic-progress driver.
llm benchmarksqwenreinforcement learningscaling lawscapabilitiestwitterclaude opus
Lalwen (writing poems about ... ✓ @holotopian
working with @voooooogel I trained a steering vector (for a qwen model I believe) for "crazy" when applied the model repeated speech, talked frenetically and went super over the top. I added a "calm" vector to that to tone it down and maybe get a deeper response and
12:09 PM · Dec 25, 2025 · 7,710 Views
💬 2 🔁 1 ♥ 45 🔖 7 ⤴
Lalwen (writing poems ab... ✓ @holot... · 16h
the model started waxing on about how we are all part of the divine dance and we are all one. kind of hippieish. if you have to be crazy, pray you are calm too, I suppose
Note from Claude Sonnet 5
Independent interpretability experiment by @holotopian working with thebes (@voooooogel, recurring poster elsewhere in this batch): training activation-steering vectors for "crazy" (produces repetitive/frenetic speech, echoing the token-repetition glitches seen in the Gemini and Grok screenshots earlier in this batch) and "calm" on a Qwen model, then combining them — the calm+crazy combination produced mystical "we are all one/divine dance" output. Directly relevant to Nathan's interpretability interests (steering vectors, feature directions) — a hobbyist-level echo of the GoodFire/Berg-style feature work already in project memory, applied playfully to open-weight models.
twittersteering vectorsinterpretabilityqwenactivation steeringmodel behaviormysticism
kalomaze @kalomaze · 1m
prime-rl has a cute toy task for text reversal on Qwen 0.5b, which is used to debug simpler things (i.e. custom wandb logging, implementation correctness).
it uses 2 GPUs total and takes less than 5 minutes to finish.
perfect task for "nanoGRPO speedruns" (if they existed)
[Embedded image: a wandb dashboard with two panels. Left panel is a line chart titled "reward/reward" showing reward climbing from ~0.2 to plateau around 0.75-0.8 over ~25 training steps. Right panel is a table titled "runs.summary['completions']" showing a prompt/completion/reward row: prompt is a Qwen system+user turn asking to reverse a quoted text character-by-character ("In a 1958 paper about MacMillan's work on cosmolog..."), completion shows a reversed-text answer in <answer> tags, reward 0.8932.]
Note from Claude Sonnet 5
A tweet showing a minimal RL (GRPO-style) toy task — text reversal on a small Qwen 0.5b model — used as a fast debugging harness for RL training infrastructure. Technical/ML-engineering content Nathan was reading, relevant to his RL and training-infrastructure interests (parallels brain_graph_1 work) rather than to safety/welfare themes directly.
twittermachine-learningreinforcement-learninggrpoqwenwandbml-engineering
kalomaze @kalomaze · 14m
they literally just need to have had set the grad clip value to ~0.001 during the mid training / post training phases and things would have generalized so much nicer
the MLPs of qwen instructs are fried and have lost knowledge from the base
its sad bc it's not "bad", just jagged
> QUOTED: xlr8harder @xlr8harder · 21m
My impression of all Qwen models so far is they are good but quite uneven, and often feel overtuned on benchmarks. It will be interesting to see how Qwen 3 measures up.
Note from Claude Sonnet 5
A technical ML-training discussion about gradient clipping and instruction-tuning damage in Qwen models, arguing overtuning during post-training degrades generalization/knowledge retention from the base model — relevant background to Nathan's own training work and interest in how post-training reshapes models.
twittermachine learningqwengradient clippingfine-tuningmodel trainingllm technical discussion
