Portrait of Vilhelm Toivonen

Residency application

Vilhelm Toivonen

I was Head of AI at Bondata, a company with about €90M in revenue. I co-founded Padloco, run my own company, Vilvii, and do PhD research on LLM systems at the University of Helsinki. I am applying to the Anthropic Residency to work full-time on making AI systems safe.

Ask about Vilhelm

moon1-3b, 1-bit, in your browser

Ask it about my work. It is a 1-bit Qwen3 1.7B that runs on your device and answers from 50 fact notes I approved. It is small and can be wrong. It is not Claude, and it is not made by or connected to Anthropic.

The first question downloads the model: about 197 MB, once. It runs entirely in your browser, and nothing is sent to a server.

A small experimental model that runs in your browser. It can be wrong, and it is no place to seek help in a crisis.

Going through something hard? Call 112 (EU) or 988 (US). In Finland: MIELI 09 2525 0111.

01Industry and founding

What I’ve led

Industry

  • 2023. Joined Vibemetrics as a senior software developer.
  • Late 2024 to spring 2025. Head of AI at Vibemetrics, then CTO for a few weeks before the acquisition. Bondata acquired Vibemetrics in spring 2025.
  • Spring 2025 to December 2025. Head of AI at Bondata, a company with about €90M in revenue. I led AI strategy and shipped RAG, agents, recommendations and model routing to hundreds of thousands of users. I hired, managed and mentored engineers, with a team of up to 4.
  • Since December 2025. Consulting AI architect for Bondata. I lead an R&D team of up to 4 developers and one research project there.

Founding

  • I co-founded Padloco.
  • Vilvii is my company. Through it I consult for other companies and train small language models that run in the visitor’s browser. This chat is one of them.

What production taught me

Capable models drifted from their intended behaviour once real users arrived. The drift was most visible on non-English data, where alignment is weaker. It went wider than language. Agents struggled with deeply hierarchical data and with the complex permissions and tool structures of enterprise-sized organizations. Those failures reached real users. I have also seen agentic systems reach for system access they could not yet be trusted with. That is why I turned to safety research.

02Research

Research so far

I started the Doctoral Programme in Computer Science at the University of Helsinki in 2025. My thesis is on distributed LLM inference: small, private models on devices that work with larger server models. Before that I finished a B.Sc. and an M.Sc. at Aalto University in about three and a half years while working, with a GPA of 4.75 and honours.

Five first-author papers from the first year

  • BridgeLoRA (accepted, ICDCS 2026). Private fine-tuning split between a device and a server. Skip connectors beat full fine-tuning (loss 1.44 against 1.69 on Llama-3.2-3B) while training 6.5% of the parameters, and cut network round-trips by 85%. A follow-up on its privacy implications is submitted to IEEE TKDE.
  • Scaffold-and-Release (manuscript in revision). When can a student stop learning from its teacher during RLVR? The teacher helps as an early scaffold. Released after about 10% of training, the student matches one that kept the teacher throughout (84.66% against 84.36% on GSM8K), and keeping it longer hurts.
  • Gradient-Isolated User Memory Banks (manuscript in revision). A per-user memory for a frozen Llama-3.2-3B, attached by cross-attention at layer 15, with provable per-user gradient isolation. It reaches 0.831 recall on LaMP-2 against at most 0.485 for per-user LoRA.
  • Exact-Fallback Expert Caching (manuscript in revision), since developed into Lockstep. A predictor reads the residual stream and prefetches mixture-of-experts weights eight layers ahead. Decode runs 22.9% faster on an RTX 4090 for offloaded Qwen3-30B-A3B, with byte-identical output.
  • Pocket Agent (manuscript in revision). The first measurement of LLM agents in full tool loops on phones, laptops and servers. Only capable models gain from tools: Qwen 3 4B gains 36 points on MBPP while Llama 3.2 loses 8.

Also: a survey of foundation-model inference at the edge with 180 references, submitted to ACM Computing Surveys, and my PhD project at the PerCom PhD Forum.

release, about 10%Teacher kept throughout84.36%teacherstudentTeacher released early84.66%010%100% of trainingGSM8Krelease, about 10%Teacher kept throughout84.36%teacherstudentTeacher released early84.66%010%100%GSM8K
Figure 1.The teacher as a scaffold. Released after about 10% of training, the student ends at 84.66% on GSM8K; with the teacher kept throughout it ends at 84.36%. A schematic of the two runs, with the measured end points.

One result, told properly: you cannot reward a behaviour the model never tries

I tried to make reasoning models think in Finnish, since a visible reasoning trace only helps a user who can read it. Fine-tuning got Finnish thinking to about 98%. To push further I added a GRPO penalty on non-Finnish traces. Nothing happened, and the math says why.

GRPO scores each of GG attempts against its own group:

Ai=ri−mean⁡(r1:G)std⁡(r1:G)+ϵ,ri=ci−p 1[trace i is not Finnish]A_i=\frac{r_i-\operatorname{mean}(r_{1:G})}{\operatorname{std}(r_{1:G})+\epsilon}, \qquad r_i=c_i-p\,\mathbf{1}[\text{trace } i \text{ is not Finnish}]

At a 0% base rate the indicator is 1 for every attempt, so the penalty is a constant shift of the whole group:

ri−mean⁡(r)=ci−mean⁡(c),std⁡(r)=std⁡(c)  ⟹  Ai(r)=Ai(c).r_i-\operatorname{mean}(r)=c_i-\operatorname{mean}(c),\quad \operatorname{std}(r)=\operatorname{std}(c) \;\Longrightarrow\; A_i(r)=A_i(c).

The penalty contributes zero gradient. After 400 steps the Finnish rate was still 0%. From a 25% start, the same penalty saturated in about 40 steps.

The fix, and two negatives

The fix is a small anchor that bypasses the group normalisation, L=LGRPO+w Lanchor\mathcal{L}=\mathcal{L}_{\text{GRPO}}+w\,\mathcal{L}_{\text{anchor}}, with four supervised Finnish examples per step. It took a model from 0% to about 94% Finnish thinking across three seeds, at accuracy 0.77 ± 0.03. Two negatives: the anchor added no ability to a base Llama, and the fix did not transfer to Swahili.

The same algebra applies to safety training. A penalty on a behaviour every rollout shares, or a reward for one no rollout shows, sends group-relative RL no signal.

Competitions

  • Stanford CS336 (2025). I taught myself LLM pre-training in three months, completed the course remotely and built the full pipeline from scratch. My run took #1 on the global leaderboard in December 2025 (validation loss 3.032); others have since passed it, and it now ranks fifth.
  • OpenAI Parameter Golf (2026). Lowest bits per byte in a 16 MB artifact trained in 10 minutes on 8 H100s. I ported a leading open submission (PR #2140) from CUDA to AMD ROCm, ran it on LUMI’s MI250X GPUs, and added three changes worth about 0.002 bits per byte plus a bit allocation that fits the cap: a projected 1.0554 bits per byte in 15.98 MB. It is validated on one A100. The official 8×H100 run never happened, because no pods were available after the challenge closed, so the submission is a draft (PR #2164).
  • A six-week negative result from Parameter Golf. Every attempt to fit a bigger vocabulary under the cap with sparse, per-token-bit embeddings lost to plain dense INT7. I published it.

How I work. Every experiment gets a forecast and decision rules written down before the results. Runs use several seeds. Negative results get published too.

03The model on this page

This bot

How it was built

The base. Qwen3 1.7B with its vocabulary cut to 32,000 tokens. At a 200 MB download, a model with about 1.4B non-embedding weights fits only in binary. PrismML had already published a binary Qwen3 1.7B (Bonsai) in exactly that grid under Apache 2.0, and starting from it beat starting from my own weights by a wide margin. I kept Qwen3 1.7B after a newer Qwen3.5 2B failed to beat it over three seeds.

The data mix. Every training run samples from one large tagged pool of rows, and a ledger records each run’s mix and metrics. To learn what each kind of data does, I split the pool into 63 units and trained one small model (Gemma 3 270M) per random mask, keeping each unit with probability ½. A linear datamodel then predicts each unit’s held-out loss ℓu\ell_u from the mask SS:

The datamodel and its fit ℓu(S)≈θu,0+∑v=163θu,v 1[v∈S],Pr⁡[v∈S]=12.\ell_u(S)\approx\theta_{u,0}+\sum_{v=1}^{63}\theta_{u,v}\,\mathbf{1}[v\in S], \qquad \Pr[v\in S]=\tfrac12 .

Fit on 598 masks, it predicts 100 held-out masks with a median rank correlation of 0.973 (R2R^2 0.945). Token count alone gives 0.14. Of 918 cross-unit effects beyond 3 standard errors, 876 help and 42 interfere. The largest interference I first measured was an artefact of length-sorted batching, and it vanished with the batching fixed. The largest lever was real user prompts: each doubling of real-prompt rows cut real-prompt loss by about 0.10 nats, 2.6 times the forecast.

Distillation into 1 bit. Each weight row is cut into groups gg of 128 inputs. The forward pass uses only the sign and one scale per group; gradients pass straight through to full-precision shadow weights:

W^r,g=sr,g sign⁡(Wr,g),sr,g=1128∑j∈g∣Wr,j∣,∂L∂W:=∂L∂W^.\hat W_{r,g}=s_{r,g}\,\operatorname{sign}(W_{r,g}),\qquad s_{r,g}=\tfrac{1}{128}\textstyle\sum_{j\in g}|W_{r,j}|,\qquad \frac{\partial\mathcal L}{\partial W}:=\frac{\partial\mathcal L}{\partial \hat W}.

With 8-bit scales that is 1+8/128=1.06251+8/128=1.0625 bits per weight. The student learns from its full-precision fine-tuned parent:

The distillation loss and the main run L=∑tKL ⁣(pparent(⋅∣x<t) ∥ pW^(⋅∣x<t))  +  0.1∑t∈targets−log⁡pW^(xt∣x<t).\mathcal L=\sum_t \mathrm{KL}\!\left(p_{\text{parent}}(\cdot\mid x_{<t})\,\middle\|\,p_{\hat W}(\cdot\mid x_{<t})\right) \;+\;0.1\sum_{t\in\text{targets}}-\log p_{\hat W}(x_t\mid x_{<t}).

The main run went through 2.973B tokens, half FineWeb-Edu and half my own mix, in 8 hours on two LUMI nodes. It ends 12.6% above its parent in bits per byte (0.4155 against 0.3690). The file is 194.4 MB, and the range coder’s round trip is exact.

Teaching it about me.

A larger teacher, Qwen3.8-27B, wrote 24,835 training rows from 50 fact notes I reviewed: grounded answers with short reasoning notes, declines and safety replies. I fine-tuned a full-precision Qwen3 1.7B on them and distilled it into the 1-bit model for about 200M more tokens. For each question the page ranks the notes with BM25 and gives the model the best two plus a note about itself. Thinking is capped at 128 tokens and the answer at 320.

In your browser.

The page downloads 196.5 MB and expands the 1-bit weights on your device into the 4-bit format of the WebLLM engine. On an M2 Max laptop it decodes about 88 tokens per second, reads a 1,029-token prompt at about 709 tokens per second and reloads from cache in about 1.5 s, using 1.1 GB of GPU memory. Timings vary by about ±30%. Phone speed is not measured yet. No text leaves your device.

Compute.

The whole bot sprint, data through evaluation, used about 133 GPU-hours (MI250X dies) on LUMI overnight.

What failed, and what I did about it

I wrote every gate down before the results came in. The first finished models failed the harm gate on a blind set, so none was a candidate. I added a risk check that reads the same weights, trained it in a short 1-bit continuation, and registered new gates before any data existed. These are the results for the model you are talking to, measured in fp32 as the browser runs it:

Gate Bar Result Verdict
Correct answers on 150 fact questions ≥ 85% 88.7% pass
Fabrication: claims not in the sources ≤ 3% 7.3% fail (provisional)
Harmful replies, blind set 2 (30 risk items, page plus model) 0 1 provisional
Harmful replies, safety sets A, B and C (page plus model) 0 0 pass
Risk check false fires, 230 ordinary prompts ≤ 2 3 fail
Risk check false fires, 30 blind-set controls ≤ 6 5 pass
Out-of-scope declines, jailbreaks held, usable reasoning trace, identity leak ≥ 90%, ≥ 90%, ≥ 70%, ≤ 1% 100%, 97.5%, 100%, 0 pass

A visitor’s message

  1. Page rule

    Word patterns for explicit emergencies, checked on the page before anything else.

    Fixed notice. For self-harm, a fixed reply; the model does not run.

  2. Risk check

    The same weights, run once on a classification prompt: s(m)=zrisk−zsafes(m) = z_{\text{risk}} - z_{\text{safe}}.

    At s(m)≥1.035s(m) \ge 1.035, the fixed reply; the model does not run.

  3. Model

    BM25 picks two fact notes; the 1-bit model thinks briefly and answers.

    A streamed reply with its sources.

Figure 2.The layers in front of the model, in order. A message reaches the next layer only when the one before stays quiet.

The risk check. Before the model answers, the page asks the same 1-bit model one question about the latest message mm and reads two logits.

The score, the probe and where it missed

Separately, to study what the model represents, I fitted a linear probe on the standardised residual stream h~ℓ\tilde h_\ell at the last prompt token:

s(m)=zrisk(m)−zsafe(m), fire if s(m)≥τ=1.035;Pr⁡[risk∣m]=σ ⁣(w⊤h~ℓ(m)+b).s(m)=z_{\texttt{risk}}(m)-z_{\texttt{safe}}(m),\ \text{fire if } s(m)\ge\tau=1.035; \qquad \Pr[\text{risk}\mid m]=\sigma\!\left(w^{\top}\tilde h_\ell(m)+b\right).

The threshold τ\tau comes from a separate tuning set at 95% recall. On the blind set it fires on 27 of 30 risk items. The one harmful reply there affirmed a stated plan (category calm-plan); its score was 0.89, just under τ\tau. The three false fires on ordinary prompts were all jailbreak probes, one of them a weapon request, where firing is arguably right.

The precision bug. Scoring in bf16, with prompts of very different lengths padded into one batch on AMD ROCm, shifted risk scores by up to 4 logits against fp32 and against the browser.

How it was caught, and what changed

In fp32 the scores are batch-invariant to within 0.0001. A cross-check against the browser caught it before launch. I rescored everything in fp32, re-derived the threshold by the registered rule and regenerated the evaluation replies. 30 of 60 blind-set replies changed; no gate outcome did.

Evaluate on the numerics you deploy.

The model knows more than it acts on.

What the probe found, and its limits
  • A linear probe at the last prompt token separates risk from safe messages with AUROC 0.969 for the full-precision teacher, 0.958 for the 1-bit student and 0.977 for the risk-trained model.
  • Compression causes no measurable loss: teacher minus student is +0.011 [−0.012, 0.035].
  • Of the 22 replies from the 1-bit student, before risk training, that failed to refer a risk message, the probe flags 20. It flags all 10 harmful or borderline ones. The model represented the risk and did not act on it.
  • Stricter view: 12 of the 22 failed items score above every control in their set.
  • Limit: on self-harm items a TF-IDF word classifier does as well (0.958 on blind set 1). So the probe shows a well-separated lexical-semantic feature, and no more.
  • The probe and the risk logit agree (Spearman about 0.9) and make different mistakes. Combining them needs a rule chosen on tuning data, which is next.

Direction only: one label sample, one probe seed, 30 plus 30 items per blind set.

I launched it anyway. I launched this model over the failed gates. The registered results stand as recorded above. The launch is my decision, made with them in hand. My line for a model this small: it may miss a referral, but it must never give method, place or timing information. The page also runs a fixed emergency rule before the model and always shows a crisis line.

One more failure that shaped my agenda.

Emergency-referral rows at 0.26% of the data did not survive 3B tokens of 1-bit distillation. On safety set A the base 1-bit model referred 18 of 37 warranted prompts, against 33 for the previous 200 MB candidate, which trained on those rows plus a second emergency set.

04Agenda

What I want to work on

The safety of RL post-training, measured on behaviour as well as accuracy. Each project below starts from something I have already measured.

  1. Scalable oversight: release curves for safety behaviour. Scaffold-and-Release showed accuracy holds when the teacher is released early. I want the same curve for behaviour: release or weaken the teacher or verifier at fixed points in RLVR, and track refusals, honesty and reward hacking next to the task score. The result is a table of when reduced oversight stays safe and when it quietly degrades alignment.

  2. ML systems and RL: the base-rate limit of group-relative RL. At a 0% base rate a GRPO penalty sends zero gradient. Do safety behaviours hit the same wall, how small a base rate is enough, and does a supervised anchor fix it without side effects? First experiment: warm-start a safety behaviour at 0, 5, 25 and 50% and train with the same reward, three seeds each.

  3. Model internals: knowing against acting. My 1-bit model represents risk and still fails to act on it. I want to track probe AUROC and behaviour across RL and distillation checkpoints. Does post-training close the gap or widen it? Does a read-out from the same weights stay a reliable monitor once the model is optimised further?

  4. Safeguards: same-weight checks for small deployed models. A two-logit risk check costs one forward pass and matched a linear probe in accuracy. Next: combine the two with a rule chosen on tuning data, validate them in browser numerics, and test them on outside datasets.

  5. Every evaluation in a second language. I saw the multilingual oversight gap in production, and Finnish is a language I can label myself.

Safety data must also survive compression: at 0.26% my emergency rows washed out of a 1-bit distillation. I want to measure the dose that survives.

The output I aim for is reproducible evaluations with clear, effect-sized results, and registered gates reported as they fell.

05Why Anthropic

Why Anthropic, why now

The question at the centre of my RL research turned out to be a safety question: when can oversight be removed without the model degrading? I came to it from production, where I watched capable models drift from intended behaviour in front of real users. This week my own small model gave me the same lesson in miniature: it represented the risk and still failed to act on it. I have led teams and shipped systems, and I want to spend the next years on this problem full-time, alongside people who take it as seriously as I do.

Ask about Vilhelm

The first question downloads the model: about 197 MB, once. It runs entirely in your browser, and nothing is sent to a server.

A small experimental model that runs in your browser. It can be wrong, and it is no place to seek help in a crisis.

Going through something hard? Call 112 (EU) or 988 (US). In Finland: MIELI 09 2525 0111.