Kevin Patrick Murphy ✓

@sirbayes · joined 19 Oct 2016

Research Scientist at Google DeepMind. Interested in Bayesian Machine Learning.

74 785Followers

753Following

1 187Posts total

433.3KViews on collected posts

Against accounts of the same size

18 posts from the last 90 days, next to the 10K–100K follower range. shown widely, but few of those viewers react.

Median views6 395this account1 384median for 10K–100K

Reach, %8.55%this account4.31%median for 10K–100K

Engagement, %0.68%this account1.92%median for 10K–100K

Metric	This account	Median for 10K–100K	Ratio
Median views per post	6 395	1 384	4.62×
Reach (views ÷ followers)	8.55%	4.31%	1.98×
Engagement rate	0.68%	1.92%	0.35×

Others in this range → Compare with another account → How these benchmarks are built →

Growth & engagement

How the posts we collected actually performed: views and reaction rate post by post, what the audience did with them, and where the follower count goes.

Views per post

5.1K12 Aug

6.8K

4.6K

3.8K

4K

4.2K

4.3K

3.6K

6K

4.4K

29K

17.4K26 Aug

19.6K31 Aug

27.8K4 Sep

Last 14 collected posts, oldest on the left. The scale is logarithmic: one post can outrun the rest a hundred times over.

Engagement rate per post

0.74%12 Aug

0.47%

0.67%

0.56%

0.66%

0.64%

0.70%

0.69%

0.89%

1.89%

0.29%

1.99%26 Aug

1.35%31 Aug

1.73%4 Sep

Reactions — likes, reposts, replies and quotes — divided by views. Median for 10K–100K accounts is 1.92%.

What the audience does

Likes51.4%2 889 in total

Reposts5.6%317 in total

Replies1.8%101 in total

Quotes0.7%40 in total

Bookmarks40.4%2 269 in total

Share of every reaction we collected for this account. Replies mean argument, reposts mean endorsement, bookmarks mean the post was worth keeping.

The follower curve appears once this account has two daily snapshots — we take one a day, and this one is on its first.

Latest posts

Image from a previous era of local optima (covering nearly all models used in statistics and ML at the time!) (Based on @ZoubinGhahrama1) https://t.co/bsgePTXIgw https://t.co/CFPXCqblqp 27.8K views · 430 likes · 48 reposts · 3 replies 04 Sep 2026 Is it just me or has CC(opus 5) become dumber recently? I got fed up with it making silly mistakes (eg forgetting to add proper logging on long-running jobs) so I decided to try out Codex... So far I am loving it: it is faster, writes *much* better prose, and great code. 19.6K views · 226 likes · 3 reposts · 33 replies 31 Aug 2026 Some updates on my "Model Discovery Agent" paper. 1) Significant new version posted to https://t.co/CUfwMhQdpP; 2) video of a talk I gave at U Toronto posted to https://t.co/y2izltTO1o ; 3) I now have a cute picture of the (polyglot) MDA workflow :) https://t.co/5VeZY78waW 17.4K views · 303 likes · 33 reposts · 9 replies 26 Aug 2026 @sirbayes You do *not* need a mechanistic model to predict the outcome of interventions or even build good theories! Vaccines predate germ theory, the 2nd law of thermo was identified by someone with the wrong model of heat, and QM doesn’t even have a standard mechanistic model 29K views · 80 likes · 2 reposts · 2 replies 12 Aug 2026 PS. I will give a talk about this work at the "RL in Big Worlds" workshop at RLC (https://t.co/BD6EfaZdjx) on 8/15 in Montreal. 4.4K views · 74 likes · 6 reposts · 3 replies 12 Aug 2026 **12/ TL;DR — three contributions** 1️⃣ **MDA**: LLM proposer + SMC + SBI + VoI, extended to the 𝓜-open regime. 2️⃣ **New SOTA** on two existing discovery benchmarks (physics, chemistry) — same accuracy, far fewer experiments. 3️⃣ **NeuronBench**: a new partially-observed, https 6K views · 44 likes · 5 reposts · 4 replies 12 Aug 2026 **10/ Novelty 3 — collapse-free *learned* summary statistics** Stochastic neurons → intractable likelihood → particle filter (accurate but slow). We instead **learn a summary statistic** (a 1-D CNN) → synthetic likelihood, **~10⁴× faster**. Key: a naive likelihood https://t.co/ 3.6K views · 23 likes · 1 reposts · 1 replies 12 Aug 2026 **11/ Why the learned summary doesn't collapse** Self-supervised encoders (JEPA-style) risk **representational collapse** (s(y)→const), patched with stop-grad/EMA hacks. We avoid it *for free*: s_φ is trained by a **supervised** objective — predict (m,θ) — which anchors it. 4.3K views · 28 likes · 1 reposts · 1 replies 12 Aug 2026 **7/ Chemistry — enzyme-kinetic rate laws (AutoSciLab)** Learn rate = f(7 controllable inputs). MDA hits its ceiling in **~8 experiments (symbolic accuracy ~56%)**; the prior SOTA (SciLab) reaches only **~42% by 60 experiments**. And MDA returns **interpretable mechanisms** — h 4.2K views · 24 likes · 2 reposts · 1 replies 12 Aug 2026 **8/ Novelty 2 — NeuronBench (a new benchmark)** Six "mystery neurons" (generalized Hodgkin–Huxley) each hide a novel ion channel that's **silent under textbook probes** — you *must* design experiments (current-clamp protocols + channel blockers) to reveal it. Unlike prior https 4K views · 24 likes · 1 reposts · 1 replies 12 Aug 2026 **9/ Biology — results** On every world, the **Bayes-forecaster beats the in-context LLM forecaster** (~10× lower error), driving forecast error down to the cell's **single-trial noise floor**. VoI and LLM-proposed designs perform similarly; both beat random. https://t.co/wxYGhm 3.8K views · 19 likes · 1 reposts · 1 replies 12 Aug 2026 **6/ Physics — the "aha moment"** On a screened (Yukawa) force, short-range launches can't tell it from a power law. Maximizing VoI, MDA designs a **long-range probe** — and the true law suddenly drops to the corner of the accuracy–complexity Pareto frontier. The model "groks" h 4.6K views · 27 likes · 3 reposts · 1 replies 12 Aug 2026 **4/ Discovery and design reinforce each other** The designed experiment identifies the mechanism the LLM proposed; the identified mechanism sharpens forecasts; sharper forecasts expose the next subtle residual → the next discovery. Result: a data-efficient discovery loop, 6.8K views · 27 likes · 3 reposts · 2 replies 12 Aug 2026 **5/ Physics — discovering force laws (DiscoverPhysics)** Infer an unknown 2-body force law from a few probe launches. MDA recovers the **exact functional form in 74%** of runs (93% numerically accurate) vs **31%/31%** for a budget-matched LLM agent — reaching (and beating) the 5.1K views · 32 likes · 5 reposts · 1 replies 12 Aug 2026 **2/ How it works** MDA couples an **LLM as a proposer** of candidate mechanisms with standard Bayesian machinery: • SMC → posterior over structure *m* & parameters *θ* (+ evidence) • SBI → intractable likelihoods • Value-of-Information → pick the next experiment Design 9.3K views · 54 likes · 2 reposts · 2 replies 12 Aug 2026 **3/ Novelty 1 — the 𝓜-open setting** What if the *true* mechanism isn't in your hypothesis set? Vanilla Bayes can only shuffle probability among the candidates you already have. MDA runs an **out-of-sample predictive check**; if the best model fails it, the LLM proposes *new* 7.5K views · 32 likes · 2 reposts · 1 replies 12 Aug 2026 Predicting the answer to interventional "what if?" questions — the outcome of an action you never took — need a *mechanistic* model, not a curve fit. And you can only learn one by *experimenting*. Experiments are costly, so the real game is **data efficiency**. Meet the Model ht 205.9K views · 1.2K likes · 173 reposts · 29 replies 12 Aug 2026 Interesting LinkedIn post from @DaphneKoller that I 100% agree with . You need to combine optimal experiment design, active data collection, AI/ML and causal modeling to make progress in bio/health. Source: https://t.co/4UsNLIyz8P https://t.co/2Z1Y3wQvus 70.2K views · 245 likes · 26 reposts · 6 replies 06 Aug 2026

Similar accounts