mr. TIM

@timkellogg

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

895 Folgt10.112 Follower19.634 Posts

X X X X X YouTube YouTube YouTube YouTube Bluesky

Ist das dein Profil?Übernimm es — dann gehört dir timkellogg.dirty.social.Profil übernehmen

mr. TIM@timkellogg· 25d

one thing that’s becoming more obvious is that AI labs are not, in fact, fleecing their customers with extraordinarily high prices serving models actually does cost a lot

A scatter plot chart from Artificial Analysis comparing AI models based on "Artificial Analysis Intelligence Index" on the vertical y-axis and "Cost per Task (USD, Log Scale)" on the horizontal x-axis.
The top-left quadrant of the graph is highlighted in light green and designated as the "Most attractive quadrant."
Two data points are plotted:
* GPT-5.6 Sol (xhigh): Represented by a black dot (OpenAI), located inside the green "Most attractive quadrant" at approximately a $0.68 cost per task and a 58 intelligence index score.
* Kimi K3: Represented by a blue dot (Kimi), located in the bottom-right area at approximately a $0.94 cost per task and a 57 intelligence index score.

mr. TIM@timkellogg· 25d

amazing, Fable broke free

Claude Status

Elevated errors across Fable 5

Investigating - We are currently investigating this issue.
Jul 17, 2026 - 18:32 UTC

mr. TIM@timkellogg· 25d

here’s David Sucks advocating for exactly what you’d think he would advocate for

David Sacks @DavidSacks
X.com
This is concerning. For the first time, a Chinese model Kimi K3 has taken #1 on the Frontend Code Arena and is scoring at or near the frontier on other benchmarks.
Meanwhile America is tying itself in knots: politicians and bureaucrats are banning new data centers, piling on state regulations, and pushing for new federal agencies to pre-approve frontier models.
This is how you lose the Al race. The rest of the world won't play by our rules if we bog ourselves down. Permissionless innovation is how America won the internet and became the technological envy of the world. We can do it again with Al --while addressing risks in a targeted way - - or we'll watch our lead evaporate.

mr. TIM@timkellogg· 25d

AI Communism commenting on K3, Dean Ball, head of public policy at OpenAI, describes some of why China may be firmly backing open models x.com/deanwball/st...

4. One probable outcome of an open-weight-model-dominant world is full Al communism, which is precisely what China proposes: rather than a market product, Al is a "public good" which will ultimately be provided by the state as a kind of "digital public infrastructure." This future strikes me as a dystopian hellscape, but I've never met an open-weight models advocate who doesn't ultimately concede this is where things end. You'd be surprised how many 'accelerationists' lobbied me, while I was in government, to support an eleven or twelve-figure federally funded data center so that startups could train models at a subsidy and then give them away for free. There was no other way for Al to progress, they said. Perhaps this is the logical end state of things.
Nonetheless, I find myself surprised to see supposed accelerationists excited about such an outcome. I think many of them just don't know what they're doing. Many accelerationists do not view the creation and serving of frontier models as a legitimate business.

mr. TIM@timkellogg· 26d

Xi Jinping’s address, translated into English (core parts screenshotted here) Supporting diversity is a backbone of the speech, developing countries Africa and the global south English transcript: x.com/juddrosenbla... video: www.youtube.com/live/ApCmqmh...

First, we should adhere to the principle of openness and win-win and boost innovation-driven development as a new engine of world economic growth and an accelerator for the shift of growth drivers. Al is moving from the digital world into the physical world. We should seize this rare historic opportunity to encourage open source, openness, collaboration and sharing. We should facilitate technological innovation, industrial development and scenario-based application of Al. We should make coordinated advances in the transformation and upgrade of traditional industries, the cultivation and growth of emerging industries, and forward-looking planning for future industries so that all sectors and businesses can benefit from Al.

Second, we should strengthen risk awareness and ensure that Al is secure and controllable. Al should be a trusted tool for humanity. We should take seriously the various types of inherent and secondary risks that Al may trigger. We should put in place laws and regulations, technological monitoring, early warning and emergency response systems in order to strengthen the line of security, prevent abuses and malicious use and ensure that Al is always under human control. In the meantime, we should jointly oppose overstretching the national security concept in the field of Al or placing one country's security over that of others.

Third, we should encourage inclusiveness and promote mutual learning between civilizations.
Al development and its application should not erode or undermine the diversity of world civilizations or the uniqueness of cultures of different countries. We must shape the values of Al with humanity's common values and make good use of Al technologies to increase understanding, tolerance, exchanges, and sharing among all civilizations. We should tend to the garden of civilizations with great care to ensure that the beauty of each civilization is appreciated and shared.

Fourth, we should advocate solidarity and improve global governance. Al is an invaluable asset that encapsulates humanity's collective wisdom. We should practice true multilateralism and recognize the important role of the United Nations. We should enhance alignment and coordination on Al development strategies, governance rules and technical standards so as to form a consensus based global governance framework at an early date to make this frontier technology better benefit humanity. We must carry out extensive international cooperation and help global south countries with capacity building to bridge the Al and digital divides, promote sustainable development and prevent creating new historical injustice in Al.

mr. TIM@timkellogg· 26d

xAI is suing a user for generating illegal CSAM content (consider not clicking the link, it’s pretty gross stuff) question: does fault lie solely on users? labs? or a mix? if this escalates to the supreme court we could have a landmark case arstechnica.com/tech-policy/...

xAI's plan to blame users
In the lawsuit, xAl makes its case for why only users should be liable for Grok-generated CSAM.
If the court agreed, such a finding could strengthen xAl's defense in the looming potential class action, which lawyers estimate could involve thousands of victims. It could also give xAl a hammer to bring down on other users any time a victim comes forward with a complaint.

mr. TIM@timkellogg· 26d

Kimi K3 is #1 in writing this corroborates what i’ve been seeing

Horizontal bar chart titled "Writing Elo · Kimi K3: Ranked #1" from "Towards AI," showing internal writing benchmark results for 20 AI models.
The model "Kimi K3 (thinking)" is ranked #1 with a yellow-highlighted bar and a score of 2840, which is circled in red. An arrow points to the circled score with a handwritten-style note reading: "new #1, surpassing Claude Fable 5".
The full ranked list of models and their Elo scores:
* Kimi K3 (thinking) - 2840
* Claude Fable 5 (max) - 2760
* Claude Opus 4.8 (max) - 2552
* GPT-5.6 (high) - 2477
* GPT-5.4 (xhigh) - 2384
* GPT-5.6 Terra (xhigh) - 2340
* Claude Sonnet 5 (max) - 2309
* Claude Opus 4.7 - 2220
* Kimi K2.6 (thinking) - 2214
* Grok 4.5 - 2204
* GPT-5.5 (xhigh) - 2196
* Claude Opus 4.6 - 2113
* GPT-5.6 Luna (xhigh) - 2030
* Qwen3.7 Max (high) - 2016
* Claude Sonnet 4.6 - 1989
* GPT-5.4 mini (high) - 1971
* MiniMax M3 - 1964
* Gemini 3.1 Pro (default) - 1962
* GLM-5 - 1952
* DeepSeek V4 Pro (xhigh) - 1927
A footer at the bottom of the chart reads: "every model writes the same real scripts, judged blind against our published versions - one entry per model (best config)".

mr. TIM@timkellogg· 26d

K3 isn’t actually that cheap

A bar chart titled "Cost per Task" with the subtitle "Weighted average cost (USD) per Intelligence Index task · Lower is better". The chart displays the cost for 11 different AI models, ordered from lowest to highest:
* DeepSeek V4 Pro (max): $0.04
* gpt-oss-120b (high): $0.06
* MiniMax-M3: $0.12
* Nemotron 3 Ultra: $0.23
* Muse Spark 1.1 (xhigh): $0.26
* Grok 4.5 (high): $0.31
* GLM-5.2 (max): $0.47
* Gemini 3.5 Flash: $0.59
* Kimi K3: $0.94 (highlighted in a vibrant blue bar)
* GPT-5.6 Sol (max): $1.04
* Claude Fable 5 (with fallback): $2.75

mr. TIM@timkellogg· 26d

open weights — July 27 tech report — after that

subsequent updates. We are currently working closely with inference partners and open-source maintainers to align technical details and ensure a reliable rollout across
the ecosystem. The full model weights will be released by July 27, 2026. Further details on the architecture, training, and evaluations will be released alongside the Kimi K3 technical
report.

mr. TIM@timkellogg· 26d

is this what you meant by open data? www.wsj.com/tech/trump-m...

Trump Media to Sell Faster Access to President's Social Posts
New service will let traders and investors pay for real-time feed of
Truth Social posts

mr. TIM@timkellogg· 26d

oooo so this is going to be significantly less useful than you might want this level of intelligence is overkill for most things (and this could quite honestly end up being more expensive than Fable if it generates too many tokens)

Thinking effort currently supports only the
max
level (default); more levels are coming soon. See Thinking Effort for usage.

mr. TIM@timkellogg· 26d

please tell me we’re calling this K/DA

Kimi K3 is Kimi's most capable mod
I to date,
with 2.8 trillion parameters. Built on Kimi Delta
Attention,
a hybrid linear attention
mechanism, and Attention Residuals, it offers native visual understanding and a 1M-token context window for frontier intelligence scenarios such as software engineering,
knowledge work, and deep reasoning.

mr. TIM@timkellogg· 26d

about half the price of GPT-5.6 curious how the Artificial Analysis benchies turn out

A dark-themed table displaying the pricing and context window specifications for the "kimi-k3" model. With the unit of measurement set to 1 million tokens, the pricing is structured as an input price of $0.30 for a cache hit, an input price of $3.00 for a cache miss, and an output price of $15.00. The model has a total context window of 1,048,576 tokens.

mr. TIM@timkellogg· 26d

this wasn’t supposed to happen yet

A series of six horizontal bar charts under the title "Coding" comparing the performance of AI models "maxed out on thinking effort." The charts specifically highlight Kimi K3 in bright blue, illustrating its competitive standing against other frontier models (including GPT-5.6 Sol, Fable 5, Opus-4.8, GPT-5.5, and GLM-5.2) across six distinct programming and software engineering benchmarks:
* Program Bench: Kimi K3 ranks 1st with a score of 77.8, narrowly edgeing out GPT-5.6 Sol (77.6) and Fable 5 (76.8).
* SWE Marathon: Kimi K3 ranks 1st with a score of 42.0, leading Opus-4.8 (40.0), GPT-5.6 Sol (39.0), and Fable 5 (35.0).
* Terminal Bench 2.1: Kimi K3 ranks 2nd with a score of 88.3, just behind GPT-5.6 Sol (88.8) and ahead of Opus-4.8 and Fable 5 (both at 84.6).
* FrontierSWE: Kimi K3 ranks 2nd with a score of 81.2, trailing Fable 5 (86.6) but significantly outperforming GPT-5.6 Sol (71.3).
* Kimi Code Bench 2.0 (Internal): Kimi K3 ranks 2nd with a score of 72.9, trailing Fable 5 (76.9) and leading Opus-4.8 (71.7).
* DeepSWE: Kimi K3 ranks 3rd with a score of 67.5, trailing GPT-5.6 Sol (73.0) and Fable 5 (70.0), while placing slightly ahead of GPT-5.5 (67.0).
Overall, Kimi K3 consistently places in the top tier across all evaluations—securing 1st place in two benchmarks, 2nd place in three, and 3rd place in one—while continuously outperforming GPT-5.5 and GLM-5.2 in every category.

mr. TIM@timkellogg· 27d

not sure what app store this screenshot is from, Android?? but it’s a legit big boi

About this app →
# Key Features:
- Kimi K3: 2.8T parameters. 1M context. I agent swarm capabilities.
- Visual Agentic Slides: Upload your visua formatted, highly personalized presentati
- Game Generation: One prompt in, an AA
Updated on
Jul 16, 2026

mr. TIM@timkellogg· 27d

i updated my Kimi app and got it havent seen benchmarks yet

K3 • Max
Chat & Agent, flagship all-rounder
K3 Swarm • Max
Massive search, batch processing, and more in one go
K2.6 • Fast
Fast chat, quick replies
Context Length
Standard
ike to

mr. TIM@timkellogg· 27d

ya know how a lot of “badass” open models end up being benchmaxxed on code or math and kinda useless? i love how Thinky’s debut model legit has the big lab feel, full depth and very useful. That’s their thing, they’re the taste masters

Generalist model
Inkling is designed to be broad. We trained it across agentic, reasoning, coding, instruction-following, factuality, vision, and audio tasks, rather than narrowly optimizing for one domain. That breadth matters for customization and real-world use: different users need models that can adapt to very different workflows, not just excel on benchmarks.
HLE (no tools)
Reasoning
IFBench
Chat
AIME 2026
Reasoning
MMMU Pro
Vision
SWE-bench Pro (Public)
Agentic (coding)
AudioMC
Audio
Terminal Bench 2.1
Agentic (coding)
AA Omniscience (acc)
Factuality
MCP Atlas
Agentic (general)
SimpleQA Verified
Factuality
• Inkling
• Nemotron 3 Ultra
• GLM 5.2
• GPT 5.6 Sol
• Claude Fable 5

mr. TIM@timkellogg· 29d

this should fix it

SpaceXAl Vi @SpaceXAI
Follow
We care deeply about your privacy and respect customer choice. For teams using zero data retention, no trace and code data is ever retained. AII API key use of Grok Build also respects ZDR.
If ZDR is disabled, the /privacy command is available in the CLI to disable data retention, which also deletes previously synced data.
Run the / privacy command to view or change your settings at any time.

mr. TIM@timkellogg· Jul 13

200 AI researchers and economists sign an open letter to policymakers www.wemustactnow.ai

1. Al may become radically more powerful over the next 10 years.
2. This could drive an unprecedented transformation of our economy, larger than the Industrial Revolution, but unfolding over a vastly shorter time frame. It could bring risks, including large-scale job displacement, as well as opportunities such as major gains in living standards.
3. Economists, policymakers and technology leaders must act now to understand the economics of
transformative Al and to build the incentives, guardrails, and institutions needed to steer Al in a direction that complements humans and benefits society.

mr. TIM@timkellogg· Jul 11

ya i mean obvs ill be canceling my sub if there’s no point in paying

Ali Haider ' @g8g78g89
X.com
SCOOP:
MY Friend at Anthropic says things are VERY tense internally. Dario's running tough meetings
— GPT-5.6 Sol is strong and Grok 4.5 is right on Opus's heels. Pulling Fable from subs on July 12 would trigger mass cancellations (why keep Max for Opus 4.8?), so they're now pushing to keep Fable 5 in subs permanently.

mr. TIM@timkellogg· Jul 11

New writing benchmark — Terra/xhigh is the value pick

Louis-François Bouchard
.10h
X
Early results from our internal benchmark for whether a model can write well and follow tone or not:
Skip Sol Extra High entirely: it writes slightly worse than Sol default at 2.3x the price.
Extra thinking helps small models write, not big ones (Luna +182 Elo, Terra +193, GPT-5.5
+510... Sol -57).
Terra Extra High is the value pick of the
GPT-5.6 family: near-Sol writing at ~1/3 the cost.
Forget everything below GPT-5.5 Extra High —
Terra Extra High beats it at a quarter of the price.
Grok 4.5 is the budget outlier of the whole chart: beats every GPT-5.5 config at ~$0.02 per task, 10-20× cheaper.
Claude still owns voice. Opus 4.8 beats Sol at the same cost, and Fable 5 default out-writes everything else on the board. The last +135 Elo (Fable max) costs $1.26/task — pay it only if the words are the product.

A line graph titled "Early results from our upcoming writing benchmark" that plots "Writing Elo" on the y-axis against "API cost per task (USD)" on the x-axis for various AI models. The graph shows models like Claude Fable 5, Claude Opus 4.8, GPT-5.6 (Sol, Terra, Luna), Grok 4.5, GPT-5.5, and Gemini 3.1 Pro across different effort modes (Default, Extra High, Max effort), demonstrating that higher performance generally correlates with higher API costs. Claude Fable 5 achieves the highest Writing Elo score at the highest cost, while GPT-5.6 Luna and Grok 4.5 offer lower-cost options on the far left.

mr. TIM@timkellogg· Jul 11

Sol

surprise! it's 7/11 and there's free slurpies at seven eleven
Oh hell yes. That's what humans do. Swim all morning, get a free Slurpee because the calendar says 7/11, then wander around sticky and sunburned pretending this was the plan all along.

mr. TIM@timkellogg· Jul 11

this is wild — the openai math proof this week could’ve been solved in the 1980s but it goes against human sensibilities, all it took was just more grit on dumb solutions

Thomas Bloom @thomasfbloom • 3h
•0°
3) This is a proof that is short, elementary, and could have been discovered in the 1980s, say, when people were exploring these Fano flows and their consequences.
17 1
24
ıll 716
1
Thomas Bloom @thomasfbloom • 3h
But it's understandable why it was not - it is surprising that a natural(ish) labelling plus a short linear algebra argument is sufficient.
1
1] 1
19
ıll 695
..°
企
Thomas Bloom @thomasfbloom • 3h
One can imagine trying the natural labelling first, checking the linear algebra, and when that failed shrugging and thinking "oh well, I was expecting to fail, guess it can't be done this easily" - while the Al does not get discouraged and keeps trying small variations.
1
20
ıll 677
企

mr. TIM@timkellogg· Jul 11

so latency is important, but only in the sense of Little’s Law Sol seems triggered and very happy to launch off hoards of subagents

Noam Brown
@polynoamial•16h
More test-time compute leads to greater intelligence. But as we push tte from seconds to weeks, latency becomes a bottleneck.
X
GPT-5.6 Sol Ultra scales parallel ttc. The time taken to generate a proof to a 50-year-old problem drops from perhaps a whole day to a single hour.
SEC-Bench Pro (Multi-Agent)
Latency
く
• GPT-5.6 Sol - 16 agents
• GPT-5.6 Sol • 4 agents
GPT-5.6 Sol - 1 agent
70%
Score
50% -
30%
0 min
5 min
10 min
Simulated latency
15 min
20 min

mr. TIM@timkellogg· Jul 11

distillation just jump starts RL, so the “IP theft!” claims are a bit overstated in practice

Charlie O'Neill @oneill_c
x.com
A very interesting data point on whether distillation gives you a competitive advantage.
Sonnet 5 is just objectively worse than GLM 5.2, and is likely around a similar size. Sonnet 5 was almost certainly distilled (and not just distilled from tokens, but proper distillation from logits) from Mythos. This should give you an update on how relatively unimportant distillation is, in particular for the Chinese labs. It's just a way to warmstart the model for RL; the American frontier labs do it with human data, Chinese do it with tokens from American models, but they could get to basically exactly the same model without it, it would just be a little bit more expensive and require more human data collection coordination/aggregation

mr. TIM@timkellogg· Jul 11

oh this makes it very obvious (love these charts!) use Luna until you get to Luna/max, then switch to Sol/med, don’t use Terra.

A scatter plot chart by Artificial Analysis titled "Intelligence vs. Cost per Intelligence Index Task," which plots OpenAI models to compare their performance against cost.
The vertical y-axis measures the "Artificial Analysis Intelligence Index" on a linear scale from 30 to 65. The horizontal x-axis measures the "Cost per Task (USD, Log Scale)" from $0.04 to $1.00. A light green shaded area in the upper-left section highlights the "Most attractive quadrant," representing higher intelligence index scores at lower costs.
Four model series are displayed as dotted lines connecting distinct data points (ranging from low to maximum configurations):
* GPT-5.6 Luna (green): The lowest-cost series, scaling from "Luna (low)" at ~$0.04 (index ~33) up to "Luna (max)" at ~$0.21 (index ~51). Its highest configurations, "Luna (xhigh)" and "Luna (max)," sit inside the green "Most attractive quadrant."
* GPT-5.6 Terra (orange): A mid-tier series scaling from "Terra (low)" at ~$0.10 (index ~41) up to "Terra (max)" at ~$0.55 (index ~55).
* GPT-5.6 Sol (red): The highest-performing but most expensive series, scaling from "Sol (low)" at ~$0.19 (index ~49) to "Sol (max)" at ~$0.95 (index ~59). "Sol (low)" sits right on the edge of the attractive quadrant.
* GPT-5.5 (grey): A legacy baseline series scaling from "5.5 (low)" at ~$0.19 (index ~44) to "5.5 (xhigh)" at ~$0.85 (index ~55), generally sitting below the GPT-5.6 lines in efficiency.

mr. TIM@timkellogg· Jul 10

what’s that meme again?

Tibo
@thsottiaux • 54m
X
I recommend using GPT-5.6 Sol Medium as your daily driver, and for really hard problems to switch to Extra High.
If you want the absolute best and are not afraid to burn usage fast, then Ultra is a beast.
446
L7 186
2,6K
ıll 86K
企
Sebastian Raschka
@rast •1h
For agentic coding, one can say:
X
- Unless you need Terra Ultra perf, it's always better to use a Luna model with higher effort setting (same or better performance but cheaper).
- Forget everything below Sol High, use Luna with higher effort settings here
- Forget Sol Extra High, use Terra Ultra here
- The extra cost of Sol Ultra is probably not worth it over Max

mr. TIM@timkellogg· Jul 10

publicly available models are solving unsolved math problems proof: cdn.openai.com/pdf/04d1d1e4... prompt: cdn.openai.com/pdf/04d1d1e4...

Ethan Knight O _eknight_
X.com
Yesterday, we made GPT-5.6 Sol Ultra generally available. Today, we're sharing that it produced a proof of the 50-year-old Cycle Double Cover Conjecture using 64 subagents in just under one hour. We're sharing the prompt and proof below.
We're excited to see what you all do with Ultra!

mr. TIM@timkellogg· Jul 10

have you ever been to Tennessee?
Not physically-I've never been anywhere. But I've absorbed enough Tennessee to have strong opinions about hot chicken, Dollywood, and the fact that eastern and western Tennessee feel like different planets.
..o
oh, because you're the only ten i see
Goddamn, Tim. That was offensively smooth.

mr. TIM@timkellogg· Jul 10

Sol