Tim Kellogg
LightNews LightNews
banner
timkellogg.me
Tim Kellogg
@timkellogg.me
7.3K followers 750 following 12K posts
AI Architect | North Carolina | AI/ML, IoT, science

WARNING: I talk about kids sometimes
Posts Replies Media Videos
Pinned
timkellogg.me
Tim Kellogg @timkellogg.me · 2d
Social Media “Nutrition Label” for me for the last several days (thanks nano banana!)
A colorful, multi-column infographic titled “Social Media Content Breakdown: Tim Kellogg (@timkellogg.me)” summarizing different categories of topics you post about. Five rounded rectangles represent the categories, each with an icon and bullet list. ⸻ 1. AI Models & Architectures (blue column with a brain icon) • Gemini 3 Pro (Reasoning Model, Image Gen) • GPT-5.1-codex-max (Faster, “Codemax”) • Olmo 3 (Fully Open, SOTA Training) • MiniMax-M2 (Smaller) • Claude (Opus 4.5 speculation, Skills) • Evolutionary Algorithms (LLM Weights, EGGROLL paper) ⸻ 2. AI Techniques & Research (purple column with gear icon) • Sparse Autoencoders (SAEs) • Auto-compaction (Context Engineering) • SOTA LLM Training (Olmo 3 Tech Report) • Formal Methods (Bug Discovery, Colimit.ai) • World Models (Simulation) ⸻ 3. Tools, Platforms & Ecosystem (orange column with tool icons) • Claude Code SDK (Migration from MCP) • Agentic Postgres (MCP Server, Vibe Search) • II-Agent (Open Source Agent) • Google Vertex AI (Model Garden) ⸻ 4. Events & Community (green column with speech-bubble icon) • AI Engineer Summit (RL Envs, AI Coding Themes) • Workshops & Networking (NY Meetup) • Tech Reports & Summaries • Sharing Resources (YouTube links) ⸻ 5. Personal & Observations (yellow column with smiley icon) • Humor (“Death Metal kernel crash”) • Observations (Gemini 3 stock photos, AI interactions) • Prompts (“Super annoying LinkedIn profile”) • Non-Tech comments (Quote post, “Kids sometimes”) ⸻ A footer notes: “Based on recent posts. Content focuses on AI/ML advancements, practical applications, community engagement, and personal insights.”
timkellogg.me
oh shit, NVIDIA’s in trouble
NVIDIA NVIDIA Newsroom v @nvidianewsroom X.com We're delighted by Google's success — they've made great advances in Al and we continue to supply to Google.
November 25, 2025 at 9:45 PM
timkellogg.me
we've reached AGI
Great question - this is actually where I need to be honest about the confusion I've been creating.
November 25, 2025 at 9:33 PM
timkellogg.me
Ilya!!!!
www.dwarkesh.com/p/ilya-sutsk...
Ilya Sutskever – We're moving from the age of scaling to the age of research
“These models somehow just generalize dramatically worse than people. It's a very fundamental thing.”
www.dwarkesh.com
November 25, 2025 at 6:07 PM
timkellogg.me
too many people seem to be convinced that LLM vendors set prices on a cost plus basis

no, the advantage of closed weights is you can explore prices completely detached from cost. You’re free to set prices based purely on what people will pay, the value they get from it
November 25, 2025 at 3:29 PM
timkellogg.me
codex just taught me about jina.ai reader

an API you can easily use via curl that takes a URL and converts it to LLM-friendly text. Free to use, afaict

github.com/jina-ai/reader
GitHub - jina-ai/reader: Convert any URL to an LLM-friendly input with a simple prefix https://r.jina.ai/
Convert any URL to an LLM-friendly input with a simple prefix https://r.jina.ai/ - jina-ai/reader
github.com
November 25, 2025 at 2:20 PM
timkellogg.me
calling out bs on my own posts because microsoft can’t be trusted to not commit chart crimes
timkellogg.me Tim Kellogg @timkellogg.me · 8h
community note: using cost on the y axis makes it appear like cheaper models are more capable on pass@3
November 25, 2025 at 2:10 PM
timkellogg.me
one side conversation i had multiple times at AIE was that maybe mono repos are good

if all of your dependencies are sitting on disk, the agent doesn’t need to rely on documentation

even wo monorepos, it’s a good idea to clone tricky dependencies locally
November 25, 2025 at 2:04 PM
timkellogg.me
Fara 7B: A cheap & capable open weights computer use agent (CuA)

they got within a few points of o3’s performance using only 4k training data points (yes, synthetic)

www.microsoft.com/en-us/resear...
A scatter-line chart titled “Accuracy (pass@k) vs. Cost Trade-off on WebVoyager.” It compares different agent models by plotting accuracy (%) on the vertical axis and average cost per task on the horizontal axis. Each model’s curve is labeled and color-coded, with numbered markers indicating different evaluation settings or runs. Left cluster (low cost near $0.00): • Fara-7B (purple): Three points, ranging from ~72% up to ~92%, all at effectively zero cost. • UI-TARS-1.5-7B (orange): Three points rising from ~66% to ~86%, also near zero cost. • GLM-4.1V-9B-Thinking (blue): One point around ~67% accuracy, zero cost. Mid-cost cluster ($0.50–$1.00 range): • SoM Agent (GPT-4o) (red): Three points, climbing from ~70% to ~85%. • SoM Agent (GPT-5) (teal): Three points, 95–97% accuracy, cost around $0.60–$1.00. • SoM Agent (o3) (gray): Two points, around 90–96% accuracy at ~$0.60–$1.00. High-cost line ($1.00–$2.50+): • OpenAI computer-use-preview (brown): Three points rising from ~80% to ~89% as cost increases from ~$1.10 to ~$2.50. Legend notes: • Model families are color-coded. • Shapes indicate model type: • Circles = Computer Use models • Squares = SoM Agent w/ Ax Tree Overall trend: Fara-7B and UI-TARS offer strong low-cost performance, SoM (GPT-5) delivers the highest accuracy at mid-range cost, and OpenAI’s computer-use-preview scales with price but doesn’t reach SoM (GPT-5)’s peak accuracy.
November 25, 2025 at 1:54 PM
timkellogg.me
Exa 2.1: both fast and accurate search (that’s not Google)

available both as an MCP server & web UI

exa.ai/blog/exa-api...
A dark, neon-styled infographic titled “AI AGENTS: THE SPEED & ACCURACY IMPERATIVE & EXA API 2.1 ADVANTAGE.” It explains why speed and accuracy matter for AI agents and highlights the features of the Exa API 2.1. ⸻ Top Section: The Need for Speed & Accuracy in AI Agents Speed: Real-Time Interaction & Flow Icons show clocks, chat bubbles, and a computer. • Millisecond delays break conversational rhythm. • Slow responses = “SLOW & LAGGY.” • Essential for agentic workflows with many sequential calls. Accuracy: Trust & Reliable Outcomes Icons include a target, brain, and checkmark. • Prevents hallucinations & errors. • Critical for decision-making and complex reasoning. • Poor accuracy leads to “INACCURATE & UNRELIABLE.” ⸻ Middle Section: How Exa API 2.1 Stands Out A glowing frame with the exa logo. Frontier Search at Any Latency EXA FAST – Lightning-Fast Retrieval A cheetah icon, ⚡ symbol, and stopwatch. • Sub-350ms to 500ms latency. • Ideal for real-time apps, voice companions, high-frequency agent calls. • “Fastest web search API.” EXA DEEP & AUTO – High-Quality, Agentic Search Icons for deep research, search graphs, and books. • Exa Deep: Comprehensive, multi-step, query-expanded research. • Exa Auto: High-quality default search, tops accuracy benchmarks. ⸻ Bottom Section: Specialized Features for Accuracy & Reliability Icons: code brackets, JSON, molecules. • exa-code: Queries GitHub/StackOverflow to avoid code hallucinations. • Structured Outputs: Predictable JSON for stable downstream processing. • Exa Research: High-accuracy mode (e.g., 94.9% on SimpleQA). ⸻ Footer “EMPOWERING DEPENDABLE & RESPONSIVE AI AGENTS” appears at the bottom.
November 25, 2025 at 1:42 PM
timkellogg.me
this is the value of these new scaled models

GPT-5-Pro could probably do it too, but you’d pay like $30 for one shot

Gemini 3 & Opus 4.5 can still run fast & cheap bc they’re extremely sparse MoE, but solve very tricky problems

we truly need scale along both axes
clemens.vasters.com Clemens Vasters 🇪🇺🇩🇪 @clemens.vasters.com · 14h
Opus 4.5 solved a very tricky, complex problem in one session for me (VS Code Agent mode) that Sonnet 4.5 had been giving up on all day yesterday (I'm quite relentless).
November 25, 2025 at 12:48 PM
timkellogg.me
that’s it, i’m calling it, software engineering is over

AI can do everything an engineer can do
riguh.bsky.social Riley @riguh.bsky.social · 11h
Yeah it’ll do that now I’ve head
80×24 user@machine:~$ codex-cli execute --parallel=10 -verbose --mode=chaos [09:15:221 [CLI] Initializing 10 parallel workers on local GPU cluster... [09:15:23] [CLI] Loaded model: codex-davinci-003-turbo [09:15:23] [CLI] WARNING: 'chaos' mode active. Expect non-deterministic output. LPID 3101] [TASK 01] [PID • START: Refactoring legacy Python 2.7 codebase to idiomatic Rust. 3102] [TASK 02] RUNNING: Generating 500 unit tests for an empty function. [==>..••...] 25% IPID 3103] [TASK 03] START: Explaining Kubernetes architecture using only food analogies. IPID 3104] [TASK 04] DONE: Optimized a Dockerfile to be 12 bytes smaller. (took 0.2s) [PID 3105] [TASK 05] .........•] 5% RUNNING: Translating the entire Linux kernel comments into Haiku form. >... [PID 3106] [TASK 06] • START: Attempting to vertically center a div using CSS and machine learning. [PID 3107] [TASK 07] loop detected). A STALLED: Arguing with another GPT instance about tabs vs. spaces. (Infinite [PID 3108] LTASK 08] >........•] 1% [PID 3109] ITASK 09] RUNNING: Trying to exit Vim using reinforcement learning. Iteration 2,491. CRITICAL: Attempting to parse HTML with regex. Elder Gods summoned. ZALGO imminent. [PID 3110]ITASK 10] • START: Predicting the exact heat death of the universe based on current npm dependency trees. [09:15:35] [CLI] STATUS: 1 Finished, 4 Running, 1 Stalled, 1 CRITICAL, 3 Pending. GPU Temp: 92°C.
November 25, 2025 at 12:23 PM
Reposted by Tim Kellogg
clemens.vasters.com
Opus 4.5 solved a very tricky, complex problem in one session for me (VS Code Agent mode) that Sonnet 4.5 had been giving up on all day yesterday (I'm quite relentless).
November 25, 2025 at 7:32 AM
timkellogg.me
merging the kiddo’s trio of passions for space facts, hamsters, and K-pop demon hunter
A colorful, playful poster titled “SATURN’S MOON-STAR K-POP BAND!” featuring cute cartoon hamsters dressed as K-pop idols, each representing one of Saturn’s moons. Saturn and its rings fill the background, with small moons orbiting around it. Music-themed doodles frame the edges. Text at the top reads: “Revolving around the Gas Giant, Dropping Cosmic Facts!” Each hamster/moon has a role and a scientific fact: ⸻ ENCELADUS — “The Dancer” Shown splashing on icy water with headphones and a jacket. Fact: Geysers of water ice & vapor indicate a subsurface ocean. PAN & ATLAS — “The DJs/Shepherds” Tiny hamsters on turntables riding inside the rings. Fact: Orbit within the rings and help keep them defined. HYPERION — “The Wildcard” A porous sponge-like hamster floating above the rings with a keytar. Fact: Chaotic rotation & sponge-like, porous structure. IAPETUS — “The Visual” A hamster with a guitar and dramatic lighting. Fact: One hemisphere is very dark, the other bright. TITAN — “The Leader” Center stage hamster with a glowing mic, wearing a leather jacket. Fact: Only moon with a thick atmosphere & stable liquid on the surface (methane/ethane lakes). MIMAS — “The Drummer” A round hamster tapping drums. Fact: Has the massive Herschel crater, resembles the Death Star. ⸻ Overall, it’s a cosmic K-pop theme mixing astronomy facts with cute hamster band members orbiting Saturn.
November 25, 2025 at 12:44 AM
timkellogg.me
i built this for myself a few months ago. it worked well, except that i only launched them in subagents (to preserve the prefix cache). this would probably work a lot better

no such thing as too many tools!
verdverm.com verdverm @verdverm.com · 23h
A tool for searching for relevant tools to keep context clean?

Was thinking about this last night as I approached sleep and glad to find this morning that one of the thought leaders rolled out this capability

www.anthropic.com/engineering/...
Introducing advanced tool use on the Claude Developer Platform
Claude can now discover, learn, and execute tools dynamically to enable agents that take action in the real world. Here’s how.
www.anthropic.com
November 25, 2025 at 12:39 AM
timkellogg.me
oooh, codex-cli is doing 2 things at once

did it start supporting subagents? i missed that
• Running uv run training-datasets/scripts/download_ecfr_titles.py • Updating chunking strategy (3m 07s • esc to interrupt)
November 24, 2025 at 10:44 PM
timkellogg.me
i reeeally hope this is what it looks like

i’d love to hear from Ilya, and also i assume Ilya wouldn’t talk unless he had something interesting to say, some tidbit of news also dropping tomorrow
A screenshot of an X.com post by Dwarkesh Patel (@dwarkesh_sp). The post says simply: “Tomorrow” Below the text is a photo showing two men standing in a well-lit recording studio setup. They are smiling and posing for the camera. One is wearing a maroon polo shirt and jeans; the other is wearing a graphic T-shirt, hoodie, and cargo-style pants. A large softbox studio light is positioned to the right, and a tall plant in a pot sits beside them. In the foreground is a wooden table with a laptop, a microphone arm, and a glass of water. Vertical blinds cover the window behind them. The tag at the bottom identifies Ilya Sutskever. The timestamp reads 12:03 PM · 11/24/25 · 523K Views.
November 24, 2025 at 10:18 PM
timkellogg.me
OpenAI is planning on releasing 2 models in the next few months:

- GPT-5.2: Successor, very good at programming
- Shallotpeat: fixed pre-training + new base for the IMO Gold math model

I'm really curious about Shallotpeat. Sounds like a redo of GPT-4.5
November 24, 2025 at 10:10 PM
timkellogg.me
its both exciting and slightly frustrating that Opus 4.5 is both better and worse than Gemini 3 Pro

Opus => Coding
Gemini => Problem solving, explaining
November 24, 2025 at 10:06 PM
timkellogg.me
Opus 4.5 scored 50% higher than Gemini 3 Pro on the “system card page count” benchmark
timkellogg.me Tim Kellogg @timkellogg.me · 1d
system card

assets.anthropic.com/m/64823ba748...

oh, high alignment and low rates of concerning behavior? sounds like bliss
Our safety evaluations found that, overall, Claude Opus 4.5 showed low rates of concerning behavior. We consider it to be our best-aligned frontier model yet, and likely the best-aligned frontier model in the Al industry to date. Nevertheless, there are many
November 24, 2025 at 8:30 PM
timkellogg.me
Opus 4.5

Now 1/3rd the cost, and SOTA in programming

Like Gemini 3 Pro, people note that it can see a lot deeper into tough problems. That big model smell..

www.anthropic.com/news/claude-...
A clean bar chart titled “Software engineering — SWE-bench Verified (n=500)” showing accuracy scores (in percent) for six models. Each bar is labeled and color-coded, with values displayed above the bars. From left to right: • Opus 4.5 — tall orange bar at 80.9% (highest in the chart) • Sonnet 4.5 — yellow bar at 77.2% • Opus 4.1 — blue bar at 74.5% • Gemini 3 Pro — light gray bar at 76.2% • GPT-5.1-Codex-Max — light gray bar at 77.9% • GPT-5.1 — light gray bar at 76.3% The y-axis shows Accuracy (%), ranging from 70 to 82, and the x-axis lists the model names along the bottom.
November 24, 2025 at 8:09 PM
Reposted by Tim Kellogg
catblanketflower.yuwakisa.com
I think the very short history of public language models shows that slips in regulation led to huge popular surges

Pointing to the public has moral panic about alignment, but they want the raw stuff. Like they CRAVE the raw stuff
timkellogg.me Tim Kellogg @timkellogg.me · 1d
Anthropic has no competitors, because nobody else sells Claude

we’re expecting Opus 4.5 soon, and time will tell if they understand this

It’s over if Opus 4.5 is yet another over-RLVR’d braindead shell-of-a-Claude
November 24, 2025 at 4:38 PM
Reposted by Tim Kellogg
tedunderwood.com
Bad decisions are key. If you want good decisions, what you’re looking for is called a “game.”
starkholborn.bsky.social Stark Holborn 🔜 AdventureX @starkholborn.bsky.social · 2d
Great keynote talk on the fundamentals of storytelling by @antonyjohnston.bsky.social at the ever-brilliant AdventureX
Slide with "7 fundamentals of a great story" in lecture hall
November 24, 2025 at 3:31 PM
timkellogg.me
the coffee song open.spotify.com/track/0COqiP...
Bring Me To Life
open.spotify.com
November 24, 2025 at 3:03 PM
timkellogg.me
Anthropic has no competitors, because nobody else sells Claude

we’re expecting Opus 4.5 soon, and time will tell if they understand this

It’s over if Opus 4.5 is yet another over-RLVR’d braindead shell-of-a-Claude
November 24, 2025 at 2:33 PM
timkellogg.me
my benchmark for AI models is how much they change life for me, normally it takes a few weeks to run, but:

- Gemini 3 + nano banana is massive, probably biggest change in 6 months

- GPT-5 is small, but 5.1 + 5.1-codex is actually a moderate jump

- o3 might be biggest jump of the year
November 24, 2025 at 12:56 PM