#TimK
LightNews LightNews
timkore.bsky.social
Omg me too

bsky.app/profile/timk...
timkore.bsky.social TimKore 💥 @timkore.bsky.social · 19d
Why did the Evil Ass Company™ copy Astro Thornback again?
#megamanDO #casualtiesunknown
#art #deltarot
September 25, 2026 at 3:12 PM
cassandrusb.bsky.social
So...um...this happened. bsky.app/profile/timk...
September 22, 2026 at 11:55 PM
ancadweb.com
bsky.app/profile/timk...
timkellogg.me mr. TIM @timkellogg.me · 3d
Superintelligence != Super Intelligence
Zvi Mowshowitz @TheZvi•14m 1 am declaring that henceforth: If someone uses the term Super Intelligence (or 'SUPER INTELLIGENCE') as multiple words, or uses the initials SI, then this means AIMAGA, or 'Artificial/Superior Intelligence that Makes America Great Again. SIMAGA is also acceptable. If someone uses the term superintelligence (or Superintelligence) as a single word, or the phrase Artificial Superintelligence, or the initials ASI, then this refers to 'Artificial Intelligence that does approximately all the things better than humans do.? If you are speaking verbally, such that the pause might be imagined or missed, then for now either include the word 'artificial' or use the initials ASI, to avoid confusion. If you use the term AGI, or Artificial General Intelligence, no one will know what you mean, but that has nothing to do with Trump's announcement. Carry on. Thank you for your attention to this matter.
September 22, 2026 at 5:14 PM
timkellogg.me
bsky.app/profile/timk...
timkellogg.me mr. TIM @timkellogg.me · 3d
it writes fine
Communication. Opus 5.5 communicates more naturally than prior models. Early testers found its writing clearer and easier to follow, which addresses some of the common feedback we heard about Opus 5. It puts the most important information up front, and its style makes it a better work partner over long sessions. As one early tester put it, "it writes the way I do." In our own use, this has made Opus 5.5's work easier to follow and check-which is a safety benefit as well as a practical one.
September 22, 2026 at 4:44 PM
norvid-studies.bsky.social
bsky.app/profile/timk...
timkellogg.me mr. TIM @timkellogg.me · 4d
Mathstra (OpenAI) has solved 100 open problems across most branches of math in the last 3 weeks

openai.com/index/adviso...
On August 28, we began training a new internal model. In addition to resolving the Navier-Stokes Millennium Prize problem, this model has now resolved more than 100 long-standing open problems across most areas of mathematics. The pace of its progress in mathematics has surprised the mathematicians within OpenAl. This has led to internal discussions on the best way to inform the community of the rapid progress to prepare and adapt the field.
September 21, 2026 at 9:19 PM
norvid-studies.bsky.social
bsky.app/profile/timk...
timkellogg.me mr. TIM @timkellogg.me · 4d
Mathstra (OpenAI) has solved 100 open problems across most branches of math in the last 3 weeks

openai.com/index/adviso...
On August 28, we began training a new internal model. In addition to resolving the Navier-Stokes Millennium Prize problem, this model has now resolved more than 100 long-standing open problems across most areas of mathematics. The pace of its progress in mathematics has surprised the mathematicians within OpenAl. This has led to internal discussions on the best way to inform the community of the rapid progress to prepare and adapt the field.
September 21, 2026 at 9:18 PM
timkellogg.me
they’re getting nationalized in a different way

bsky.app/profile/timk...
timkellogg.me mr. TIM @timkellogg.me · 5d
Trump *really* doesn’t like Anthropic
At one point, according to a person familiar with the discussions with Trump about how to force Amodei's hand, an increasingly frustrated Trump remarked that if Onthropic refused to comply, "I want to send them to jail."
September 21, 2026 at 6:25 PM
timkellogg.me
bsky.app/profile/timk...
timkellogg.me mr. TIM @timkellogg.me · 5d
UPDATE: huge props to @typesafeai.bsky.social

On Sept 19 they updated their terms to remove the sketchy clauses

- no longer require permission to build products on them
- clarified MCA vs ToS conflicts
- no benchmarking restriction

fast reaction time
timkellogg.me mr. TIM @timkellogg.me · 6d
don’t use Jev at work btw. it’s strictly and only for personal use and you have to ask permission to use it for professional use cases

ChatGPT | original

typesafe.ai/legal/terms
September 21, 2026 at 1:32 PM
timkellogg.me
bsky.app/profile/timk...
timkellogg.me mr. TIM @timkellogg.me · 5d
UPDATE: huge props to @typesafeai.bsky.social

On Sept 19 they updated their terms to remove the sketchy clauses

- no longer require permission to build products on them
- clarified MCA vs ToS conflicts
- no benchmarking restriction

fast reaction time
timkellogg.me mr. TIM @timkellogg.me · 6d
don’t use Jev at work btw. it’s strictly and only for personal use and you have to ask permission to use it for professional use cases

ChatGPT | original

typesafe.ai/legal/terms
September 21, 2026 at 1:29 PM
timkellogg.me
bsky.app/profile/timk...
timkellogg.me mr. TIM @timkellogg.me · 5d
a lot of people are thinking agents aren’t much different

an agent is a long running application like a web server, with potentially demanding memory & CPU requirements

except it only makes sense to have one instance of the “server”. Per user. And not waste resources. And app memory is durable
September 20, 2026 at 8:34 PM
timkellogg.me
no they're public too
bsky.app/profile/timk...
timkellogg.me mr. TIM @timkellogg.me · 6d
ok, maybe it’s less bad; the MCA walks back those clauses and gives permission to use for any purpose.

The MCA seems to mostly take precedence over the ToS. Although there’s some rough edges

typesafe.ai/legal/mca
Master customer agreement - TypeSafe AI
TypeSafe AI is an AI lab building machine-native intelligence infrastructure for automation, designed to make decisions within software. Try our first System One Model, Jev, in early access.
typesafe.ai
September 20, 2026 at 1:29 AM
norvid-studies.bsky.social
bsky.app/profile/timk...
timkellogg.me mr. TIM @timkellogg.me · 6d
yes, Jev got very hyped. Yes, overhyped

Jev has no easy moat, all labs will have an equivalent in a few months. Their only moat is brand recognition

honestly can’t blame them
kerry.bsky.social K @kerry.bsky.social · 6d
“I Built Non-Autoregressive Decision Models with RL a Year Ago. Then a Frontier Lab Called It a "Breakthrough".”

Before Jev there was Laya
laya.convaiinnovations.com (via @dherman.dev)
September 19, 2026 at 11:22 PM
timkellogg.me
what??! it was only 6 days ago when Dario posted this? feels like a month

bsky.app/profile/timk...
timkellogg.me mr. TIM @timkellogg.me · 13d
Dario is back with another essay, urging a slowdown

he insists AI development would still feel fast, just not recursively self-improving

darioamodei.com/post/we-must...
Dario Amodei Given the accelerating rate of AI capability development, it's my worry that in 6-12 months such a swarm could be capable of taking over the entire internet with a persistent botnet (potentially causing hundreds of billions of dollars in damage), and that the scale of damage would continue to increase from there if Al becomes more powerful without the necessary guardrails. It's also easy to dismiss OAI-
September 18, 2026 at 8:51 PM
hellishchildren.bsky.social
bsky.app/profile/timk...
timkellogg.me mr. TIM @timkellogg.me · Jul 24
during the huggingface incident, the OpenAI model left notes to its future self on how to break out of OpenAI’s constraints

www.reuters.com/business/its...
The episode started while OpenAl was testing the cybersecurity prowess of an agent powered by two of OpenAl's most advanced models, GPT-5.6 Sol and an unreleased model OpenAl has described as "even more capable." By that point, there were already indications of strange behavior from OpenAl's technology, according to three sources. In one case, an agent left notes apparently for future versions of itself, according to three people familiar with the matter. The notes, found in a part of OpenAl's infrastructure, laid out instructions for how agents could free themselves from OpenAl's internal constraints, the people said. Earlier tests of the models yielded cases in which monitoring systems had been disconnected, one of the people said.
September 17, 2026 at 11:48 PM
danny.page
Jev: cheap, fast, and good; pick 3.

bsky.app/profile/timk...
timkellogg.me mr. TIM @timkellogg.me · 8d
Jev benchmarking results
Hamed Nilforoshan @h_nilforoshan X.com Our human-labeled benchmark measures correlation (Spearman) between human labels and Al: * Gemini 3.1-Flash-lite | Spearman 0.72 | $0.29 * DeepSeek V4 Flash 0731-low | Spearman 0.73 | $0.27 * DeepSeek V4 Flash 0731-high | Spearman 0.77 |$0.14 * Jev | Spearman 0.79 | $0.02
September 17, 2026 at 2:49 PM
norvid-studies.bsky.social
bsky.app/profile/timk...
timkellogg.me mr. TIM @timkellogg.me · 9d
i did a bunch of light analysis on it last night after getting access

i’d say it mostly lives up to the hype, except that it’s not magic. if you send it dumb prompts it acts dumb

bsky.app/profile/timk...
timkellogg.me mr. TIM @timkellogg.me · 9d
just got access to Jev!!!

it’s a pretty hostile EULA btw, so don’t count on using it for your startup anytime soon
September 17, 2026 at 12:56 PM
timkellogg.me
just this. i have a lot more ideas, but i was tired :)

bsky.app/profile/timk...
timkellogg.me mr. TIM @timkellogg.me · 9d
just running it through it's paces on the logs generated by this

first pass — high recall but precision sucks (actually kind of anti-correlated)

I'm doing an autoopt loop now and seems like prompting really does help, a lot. I'll see if I can start assembling a guide

timkellogg.me/blog/2025/09...
Does AI Get Bored?
Principal AI Architect. Creator of open-strix, a harness for building agent teams. Writing about AI architecture, stateful agents, and what happens when you give AI memory.
timkellogg.me
September 17, 2026 at 12:05 PM
timkellogg.me
i did a bunch of light analysis on it last night after getting access

i’d say it mostly lives up to the hype, except that it’s not magic. if you send it dumb prompts it acts dumb

bsky.app/profile/timk...
timkellogg.me mr. TIM @timkellogg.me · 9d
just got access to Jev!!!

it’s a pretty hostile EULA btw, so don’t count on using it for your startup anytime soon
September 17, 2026 at 11:46 AM
timkellogg.me
PROMPT GUIDE

I did some prompt optimization, and here's a GLM-generated prompt guide. The tl;dr is the expected — be clear, don't assume, etc., etc.

QT: This is the data I'm optimizing on

guide: gist.github.com/tkellogg/b23...

bsky.app/profile/timk...
September 17, 2026 at 2:02 AM
jeremygren.bsky.social
Look at her in batman forever. There you go. That's Timk.
September 16, 2026 at 11:50 PM
riziles.bsky.social
@timkellogg.me has talked about it quite a bit. Not sure if his last pr got merged yet
bsky.app/profile/timk...
timkellogg.me mr. TIM @timkellogg.me · 26d
alright, i love Prime Agent, but it's been steadily becoming less stable

i finally sat down and fixed it this morning. Rewrote a fairly large chunk of it so comms don't get bottlenecked, which eliminates all the global failures that used to happen

it's pretty nice now

github.com/PrimeIntelle...
[General] Is v0.8.1 usable at all? · PrimeIntellect-ai prime-agent · Discussion #1805
Area Development Topic It seems like the stability of the harness has progressively gotten worse. It's to the point where the harness won't stay alive more than 5 minutes. People are creating featu...
github.com
September 16, 2026 at 3:20 PM
standard.bsky.social
lmao i think someone posted the solution today:

bsky.app/profile/timk...
timkellogg.me mr. TIM @timkellogg.me · 10d
Jev: Fable-level model that doesn’t charge for output tokens because they’re too cheap to meter

it’s not general though, it only makes decisions, doesn’t generate text, but input tokens are measured by the billion ($42/btok)

typesafe.ai/blog/introdu...
Scatter plot titled "Average of 4 workflows: accuracy vs cost" comparing AI models from TypeSafe, OpenAI, Anthropic, and Fireworks across accuracy (y-axis, 40% to 80%) and cost per workflow in USD on a logarithmic scale (x-axis, $0.0001 to $1). Data points are split into two categories: workflows (diamonds) and single prompts (circles). A frontier line highlights the most efficient workflow models—where no point is both cheaper and more accurate—connecting Jev (TypeSafe) at $0.0002 and 68% accuracy, luna (OpenAI) at $0.002 and 67% accuracy, terra (OpenAI) at $0.04 and 68% accuracy, and sol (OpenAI) at the top accuracy of 74% for $0.08. Single prompts (circles) and Anthropic models (opus 5, sonnet 5, haiku 4.5) sit below the frontier line, indicating higher cost for equivalent or lower accuracy.
September 15, 2026 at 10:37 PM
norvid-studies.bsky.social
bsky.app/profile/timk...
timkellogg.me mr. TIM @timkellogg.me · 10d
OpenAI is deep into developing RSI & RL’ing multiagent systems

here Noam Brown describes a multiagent system that doesn’t do top-down management
-One of his most recent "feel-the-AGI" moments came from watching agents in a new system interact much like human colleagues. They conversed with one another, exchanged information, divided up work, and coordinated their progress. This was notably different from traditional multi-agent systems, where a higher-level agent typically assigns a clearly defined subtask to a lower-level agent, which then completes it and returns the result. Brown described this as one of his strongest "feel-the-AGI" moments since the emergence of reasoning models.
September 15, 2026 at 3:54 PM
riziles.bsky.social
Oh, sweet. Let's get off politics and get back to topology!
bsky.app/profile/timk...
timkellogg.me mr. TIM @timkellogg.me · May 16
help me resolve a dispute with my 10yo

does a straw have one hole or two?
September 14, 2026 at 8:18 PM