sam ehrlich

Driveline Baseball Ops Analyst, MIZ M.S. Data Science 2023
CHI Cubs and Bears

[Driveline Baseball @DrivelineBB]

.@AFredericks2027 was already hitting .450 with a .702 SLG, 4 triples, and 2 home runs, and committed to Nevada.

But she still wanted more.

After 7 weeks of training at Driveline:
Max EV: 73.2 ➡️ 77.6 mph (+4.4)
Avg EV: 64.0 ➡️ 66.6 mph (+2.6)
Bat Speed: 58.5 ➡️ 60.5 mph (+2.0)
Max Distance: 241 ➡️ 261 ft (+20)

[Tomas Nelson @tomashnelson]

Quickest and Slowest to the ball in 2026 (300+ swings)

Time to contact = 1.3636 × swing length / bat speed

Swing acceleration = 0.03343 × bat speed² / swing length

What do you see?

[【公式】ドライブライン ベースボール ジャパン Driveline Baseball Japan @DLBB_JP]

⚾️ バットスピードが速ければ、良い打撃ができるのか？

実は、速さだけでは説明できない打撃の違いがあります。

ドライブラインベースボールジャパンの最新noteでは、バットの動きを3次元で分析する「バットパス最適化モデル」を詳しく解説！

📊 インパクト時のデータから18点のバット軌道を推定
📊 パワーとコンタクトのバランスを分析
📊 打者一人ひとりに合った改善方法を探る

鈴木誠也選手やムーキー・ベッツ選手などの分析例に加え、理想の軌道を身につけるためのトレーニング方法も紹介しています。

「速く振る」だけではない、科学的な打撃改善の世界をぜひご覧ください！⚾️

📖 記事はこちら
https://t.co/eSg4TT6TGt

[Deepak Pathak @deepakpathak]

Scalable simulation for robotics is here.

Coding agents can now generate environments at scale. The next bottleneck? Simulating them at scale.

Introducing Steel: a scalable physics solver for heterogeneous environments. Our first open-source release at @SkildAI, built in partnership with @NVIDIA.

Our ambition is to rally the industry around closing the sim2real gap. Contributions welcome from you, your organizations, and your agents! ❤️

[Nick Pollack @PitcherList]

Had a great time with @blandalytics getting a tour and talking shop at @DrivelineBB!

No shock to see that they are working on some awesome projects.

Glad I can finally start my plyos routine now

[Kyle Boddy @drivelinekyle]

Emily (@CoolbaughDL) will be in Las Vegas at the MLB General Manager meetings this week for all things Driveline + Dominican Republic.

I'll be in tagging in via Zoom as needed! Schedule a meeting via email: [email protected].

https://t.co/YXp8hyinKQ

[Tomas Nelson @tomashnelson]

Hitters with most Run Value to gain from path optimization

Closest to average path ➡️ Closest to optimal path

What stands out🧐

@SamEhrlich

[andrew le @andrewletics]

Introducing Launchpad Dash.

[Driveline Baseball Dominican Reublic @DrivelineBBDR]

For immediate release:

Driveline Baseball is opening its first Dominican Republic facility, in partnership with Brian Mejia of the Dominican Prospect League. Doors open November 16 in Guerra.

[Driveline Baseball @DrivelineBB]

Trajekt is one of the most beneficial tools in the game to simulate game-like reps, check out AZ Hitting Coordinator Jacob Hirsh breaking down how we use it with our athletes at Driveline! Full video on our Youtube.

[Terence Liu @terencezliu]

Discovered RF-DETR through @roboflow's Visual Intelligence Summit. I often see GPU batch=1 latency numbers for real-time use, but not large-batch chip-saturating throughput numbers for offline batch-processing. So opening a thread here.

Short intro with YOLO26. YOLO26 is convolution-based, and runs fast on modern CPUs, but AGPLv3 license. RF-DETR is transformer-based (DINOv2 backbone), slower on CPUs, Apache 2.0. Neither uses NMS for post-processing anymore.

On an AWS c8i instance (AMX acceleration) with OpenVINO, single-threaded 1 image inference:

YOLO26 small at 640x640: 42ms
RF-DETR nano at 384x384 (out-of-box comparable): 61ms

But on NVIDIA GPUs with TensorRT, batch=1 latencies compare similarly, as shown in official benchmarks. RF-DETR generalizes very well for real-world specific objects due to its transformer arch.

https://t.co/70QUhArltN
https://t.co/g5uZ82HaF0
https://t.co/AWYoNhc2Mf

[Driveline Baseball @DrivelineBB]

Austin Riley has gone from one of the best hitters in the league to producing at a league-average level.

‘21-‘23: 4.9 WAR/600, .382 wOBA
‘24-‘26: 2.0 WAR/600, .319 wOBA

So what changed?

His contact quality has dipped in his down years, especially on fly balls. He’s started hitting the ball in the air more: his average launch angle went from 13.6 to 17.1 degrees. But that contact in the air is much less productive. On balls hit between 25 and 50 degrees, his home run rate fell from 26% in ‘21-‘23 to 17% in ‘24-‘26. On line drives, he’s been similarly productive. His top-end power has slipped too, with max exit velo down from 116.1 to 114.7 mph.

The damage shows up everywhere. His K% went from 24.5% in ‘21-‘23 to 28.9% in ‘24-‘26, and his walk rate dropped from 8.1% to 7.2%. This season it’s been at its worst. His xwOBA fell from the 94th percentile in 2023 to the 51st in 2026, his whiff rate climbed from 27.8% to 31.9%, and his K% is now 31.9%.

He’s missing more offspeed pitches and breaking balls early, and that’s where a lot of those whiffs are coming from.

Part of it is the swing itself. Compared to the second half of 2023:

Attack angle: 9.98 to 11.41
Horizontal attack angle: 3.64 to 5.35
Tilt: 34.04 to 31.29 (flatter)
Point of contact: 1.72 ft to 1.91 ft (further out front)

These changes seem to be the result of compounding compensations, starting with his intercostal strain in early 2024. He likely had to move his point of contact out front in ‘25 and ‘26 to make up for a dip in bat speed in ‘24 (76.9 to 75.8 mph).

Riley would likely benefit from a swing adjustment that steepens his tilt and shifts his path back deeper in the zone. Those changes could be exactly what he needs to get back to his ‘21-‘23 production.

-@Stevie_T12

[sam ehrlich @SamEhrlich]

@tomdoyo Roughly 4 years on the $20 plan or 10 months on the $100 plan. Again, I think owning is better than renting for many reasons if you can afford the upfront cost.

[sam ehrlich @SamEhrlich]

Really cool use case for self hosted agents. Having an agent I can host and allow my parents to use as they get older and have questions is a great idea

[andrew le @andrewletics]

Just copped an M6 Mac mini 32gb/256gb. After using GrokBot/Hermes Agent Mobile (shoutout @NousResearch), I realized my parents (non native English speakers) rely on me a lot- and that these assistants honestly would be most useful to them.

[sam ehrlich @SamEhrlich]

@tomdoyo Self host would allow full end to end control of the model and harness and a single time payment for the hardware. Great for tasks that are not requiring the highest level frontier model outputs to still get the results needed like answering questions for foreign parents

[Tyler Preece @Tyler_Preece_3]

PCA was top 5 in xwOBA gain vs baseline in June, August AND September

Dug into his numbers and a few things jumped out

Chase 41.7% ➡️ 34.2%
BB 4.5% ➡️ 11.4%
Bat speed 72.7 ➡️ 75.1

One of my favorite players to watch

Data via Samcast by @SamEhrlich

[Alex A - mirrash.eth 🦇🔊 @mirrash7]

Adjusted my sports pipeline for all sports!

-Every player segmented, tracked + identified
-Rebuilt as 3D bodies on a calibrated court/field
-Ball, puck + disc tracked in 3D
-Replay from any angle

Here it is for volleyball, hockey, pickleball, basketball, soccer

All from a single broadcast camera!!

[Kyle Boddy @drivelinekyle]

Amazing to watch the growth over the last year of Driveline. Operations in Japan, Dominican Republic, and stealth deployments under NDA are going super well. Launchpads being installed in colleges nationwide!

Pro athletes streaming into all locations.

Alumni dominating upper echelons of the game.

Several research papers in the final stages of peer review, a patent recently granted for our Intended Zone Projector technology...

And some of our greatest R&D / Sports Science accomplishments haven't even been finalized/announced yet.

Much more to come as we continue to pick up velocity of development and shipping. Right in the face of the doomers and haters, which provides me a non-zero amount of joy, I must admit.

[Jeremy Frank @MLBRandomStats]

Kanz is an outstanding leader and one of the brightest and most ambitious people I’ve gotten to work with. Cannot say enough good things about him.

Mariners fans are going to love him

[sam ehrlich @SamEhrlich]

CV for counting and anomaly/defect detection on assembly lines seems to be the most popular use case that I have seen lately.

However, the trailing keypoints on the fish as they fly across the screen is hilarious

[Sadao Tokuyama @tokufxug]

手のひらサイズの小型デバイスで大量に流れてくる魚を毎秒50フレーム検知と認識！🐟

🔹顧客の依頼を受け、魚の認識・カウントシステムを開発
🔹98%は魚を数えられた割合。検出率・追跡精度ではない

使用したハードウェアや評価条件は未公表
開発：Oleksandr Melokhin氏
情報元はリプ欄。