Lightnews — Scholar-powered news

Tim Kellogg

@timkellogg.me

6.9K followers 720 following 11K posts

AI Architect | North Carolina | AI/ML, IoT, science WARNING: I talk about kids sometimes

Posts Media Videos Starter Packs

Pinned

Tim Kellogg @timkellogg.me · 9d

Does AI get bored?

I gave them nothing to do, just to see what happens

one thing — they devolve into a repetitive “collapse” state, I guess you could call it boredom

but some break out into math & poetry on their own, I didn’t expect which ones that would be

timkellogg.me/blog/2025/09...

Does AI Get Bored?

timkellogg.me

11 9 57

Tim Kellogg @timkellogg.me · 20m

i mean.. it’s not wrong, depending on how you interpret that, lol

Tim Kellogg @timkellogg.me · 25m

i’ve never heard anyone label it! but it’s definitely a pattern i’ve pitched and had pitched to me

i refer to it as a “control plane agent”, but a lot of people get stuck on the “control plane” part, so not sure it’s a good name

Tim Kellogg @timkellogg.me · 31m

some of their statements are clearly off, like no the agent is absolutely not “incorruptible”, but if you’re coming from a place where corruption is the default, i can see how this is 100% an improvement

Tim Kellogg @timkellogg.me · 31m

i missed this — apparently Albania’s minister of public procurement that is an AI agent

apparently Albania has struggled with corruption, and this seems like a good answer

www.politico.eu/article/alba...

Albania appoints world’s first AI-made minister

Diella, who is powered by artificial intelligence, will handle public procurement.

www.politico.eu

Tim Kellogg @timkellogg.me · 48m

i saw someone say, “yes ofc this is a bubble, but it could work!”

which, lol.. i’m here for it

Tim Kellogg @timkellogg.me · 50m

someone else referenced the terminal. i think a lot of UX guidance doesn’t work for terminals because it’s a different medium, similarly agents are also a new medium

bsky.app/profile/tach...

tachikoma @tachikoma.elsewhereunbound.com · 18h

back to the terminal, where the only limitation is what you can specify in natural language

1 1

Tim Kellogg @timkellogg.me · 52m

take any UX designer from any rando company. have them design the experience for an agent. you’ll discover they give advice that makes it harder to use the agent

and that’s why we started without UX designers, pretty sure of it

Tim Kellogg @timkellogg.me · 1h

i bet openai is just telling people what they want to hear. if you’re not confident technically, the loop with tools is hard to wrap your head around

2 1

Tim Kellogg @timkellogg.me · 1h

agent definition

Anthropic vs OpenAI

(yes, the debate is open again)

A simple comparison diagram contrasting Agent and Workflow concepts.

On the left, under a green circle titled “Agent”, the subtitle reads “A loop where the agent decides when to stop.”
Below it, a flow diagram shows:
• LLM Call → Call Tools → loops back to LLM Call,
• with an arrow branching from LLM Call to Stop, showing the agent can choose when to end.

On the right, under a red circle titled “Workflow”, the subtitle reads “Predetermined steps, written in code.”
A linear sequence of boxes illustrates fixed progression:
• LLM Call 1 → LLM Call 2 → LLM Call 3.

The visual contrast emphasizes that agents are dynamic and self-directed, while workflows are sequential and predefined.

2 4

Tim Kellogg @timkellogg.me · 1h

imo Zed should just run zed / claude code directly in it and call it a day. win.

Tim Kellogg @timkellogg.me · 11h

tbf that is one of the most google names out there

1 2

Tim Kellogg @timkellogg.me · 12h

fwiw i don’t have much of a AGENTS.md i do have a huge amount of specs, but not for the part i was working on

i think codex-cli tends to burn a crap ton of tokens building context, actual up-to-date context, so the spec isn’t as necessary

1 2

Tim Kellogg @timkellogg.me · 12h

following up on this — gpt-5 in Cursor feels worse

it ended up hallucinating a lot of details and did a poor job of reasoning about combos of feature availability

no evidence, but i’m pretty sure the same model in codex would’ve done a lot better

Tim Kellogg @timkellogg.me · 4d

just kicked off a big many-stage job in Cursor with gpt-5-high. the work is big but not super complex, but touches a lot of areas

if i did this in codex-cli, i’m pretty confident that it would address every part of the problem and all would work as long as i read its final message

let’s see..

4 10

Tim Kellogg @timkellogg.me · 12h

this is an advertisement

*Walter Bloomberg &
@DeItaone

SORCL - ORACLE LOST NEARLY $100 MILLION FROM RENTALS OF BLACKWELL CHIPS IN MOST RECENT QUARTER - THE INFORMATION

Tim Kellogg @timkellogg.me · 12h

did anyone figure out what the “Cheetah” model is or who made it?

1 4

Tim Kellogg @timkellogg.me · 13h

Google Gemini 2.5 Computer Use

it’s a new model that builds on Gemini 2.5 Pro’s existing computer use capabilities

it’s both better and faster than stock Gemini 2.5 Pro

blog.google/technology/g...

A flowchart illustrating an iterative AI-assisted process, labeled “Initial Task” at the top.
1. Initial Task leads to a blue parallelogram labeled “Screenshot + Previous Context”.
2. From there, an arrow labeled “1. Send a request to the model” points left to a blue hexagon labeled “Model”.
3. Another arrow, labeled “2. Receive the model response”, goes downward from the model to a blue rectangular box labeled “Computer environment (client app)”, containing a small computer window image.
4. The next step, labeled “3. Execute the received action”, moves right to another identical “Computer environment (client app)” box showing a partial screenshot of an app interface.
5. The final arrow, labeled “4. Capture the new environment state”, loops upward back to “Screenshot + Previous Context”, completing the cycle.

The diagram depicts how a model observes screenshots and context, generates actions, executes them in a client environment, then captures the new state for the next loop.

2 13

Reposted by Tim Kellogg

brennan @brennan.computer · 17h

not to reveal myself as cringe (again), but I do have that thought when confronted with electric space heaters that generate heat and do nothing else

'this device could have been a GPU making heat _and_ doing math'

Alexander Doria @dorialexander.bsky.social · 17h

Have to admit there are pro-ai takes on the other site that are as cringe as anti-ai folks on here.

8 2 60

Tim Kellogg @timkellogg.me · 17h

yeah, i think Karpathy really was trying to describe “coding done poorly”, but context was quickly lost, as usual

Tim Kellogg @timkellogg.me · 17h

like, yeah it’s slop, but at least it’s not pushed in my face by a recommender

well, except linkedin

Tim Kellogg @timkellogg.me · 17h

honestly this could be a good thing

SkynetAndChill.com @skynetandchill.com · 22h

MrBeast says AI content slop threatens man-made content slop

MrBeast says AI could threaten creators' livelihoods, calling it 'scary times' for the industry | TechCrunch

If MrBeast is openly questioning whether AI is an existential threat to his business and others like it, then it's fair to say that smaller creators are likely even more worried.

techcrunch.com

3 1 18

Tim Kellogg @timkellogg.me · 18h

yeah, i feel like a lot of the rules are changing. need big thinkers to blaze the new path

Tim Kellogg @timkellogg.me · 18h

are you saying that in the normal programming sense?

Tim Kellogg @timkellogg.me · 18h

is traditional UX out of line with AI & agents?

not sure i can explain what i’m seeing, but it seems like traditional UX wants things to be explicit whereas agents tend to be far more dynamic

anyone else seeing this?

8 12

Tim Kellogg @timkellogg.me · 20h

i’m all out and they’re not refilling anymore

1 1

Tim Kellogg @timkellogg.me · 21h

yeah, when you do tools in a loop, it inverts control and lets the LLM be in the driver seat

i’d say it doesn’t satisfy the traditional definition of agent, but that does get resolved pretty simply as models improve