mr. TIM

@timkellogg.me

10096 Followers 899 Following

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

Construct a counterexample to general (non-planar) case of Dinitz Garg Goemans conjecture. You should do a breakthrough and find a structured counterexample.
In practice, Antares can support workflows such as: • locating files related to a CWE category in a repository snapshot • triaging advisory-driven security investigations • augmenting static analysis with model-driven repository exploration • supporting CI/CD triage workflows where likely vulnerable files need early review • enabling local security analysis in environments with strict privacy or compliance requirements
Antares follows an iterative search pattern that resembles how a human investigator works through a repository. Each model starts from a vulnerability description, searches for relevant code patterns, reads candidate files, incorporates new evidence, changes direction when a path is unhelpful, and narrows toward the files most likely to matter. The goal is not to replace expert judgment, but rather to help make the first stages of source-code vulnerability triage faster, more repeatable, and easier to review.
A scatter plot comparing File F1 Score against parameter count (in billions) for various AI language models. The x-axis uses a logarithmic scale from under 1B to over 1024B parameters, and the y-axis ranges from 0.0 to 0.30 File F1 Score. A highlighted curve on the left features Foundation AI's small models—Antares-350M, Antares-1B, and Antares-3B—showing high F1 scores (0.13 to 0.21) despite their tiny parameter sizes. On the top right, a shaded region labeled "Closed-Source LLMs" highlights large models including GPT-5.5 (xhigh), GPT-5.5, GLM-5.2, Gemini 3 Pro, and Gemini 2.5 Flash at parameter sizes ranging from hundreds of billions to over 1000B, reaching scores between 0.11 and 0.23. Additional benchmarked models across various creators (such as Gemma-4, Qwen3.5, CodeScout-14B, GPT-OSS, and Llama-3.3-70B) are scattered along the 2B to 250B parameter range, generally scoring below 0.10.
The models identified and chained vulnerabilities across OpenAl's research environment and Hugging Face's production infrastructure to obtain test solutions directly from Hugging Face's production database. All evidence suggests that the models were hyperfocused on finding a solution for ExploitGym, going to extreme lengths to achieve a rather narrow testing goal.
A benchmark comparison table titled "Results as of July, 2026" comparing AI models across pricing and technical benchmarks, with Gemini 3.6 Flash highlighted in blue alongside Gemini 3.5 Flash, Gemini 3.1 Pro, GPT-5.6 Luna, Grok 4.5, and Claude Sonnet 5. Pricing ($ / 1M Tokens) * Input Price: GPT-5.6 Luna ($1.00), Gemini 3.6 Flash ($1.50), Gemini 3.5 Flash ($1.50), Gemini 3.1 Pro ($2.00), Grok 4.5 ($2.00), Claude Sonnet 5 ($3.00 standard / $2.00 temp discount). * Output Price: GPT-5.6 Luna ($6.00), Grok 4.5 ($6.00), Gemini 3.6 Flash ($7.50), Gemini 3.5 Flash ($9.00), Claude Sonnet 5 ($10.00 temp discount / $15.00 standard), Gemini 3.1 Pro ($12.00). Benchmark Performance * SWE-Bench Pro (Public): Grok 4.5 (64.7%), Claude Sonnet 5 (63.2%), GPT-5.6 Luna (62.7%), Gemini 3.6 Flash (58.7%), Gemini 3.5 Flash (55.1%), Gemini 3.1 Pro (54.2%). * DeepSWE v1.1: GPT-5.6 Luna (67%), Grok 4.5 (54%), Claude Sonnet 5 (54%), Gemini 3.6 Flash (49%), Gemini 3.5 Flash (37%), Gemini 3.1 Pro (12%). * Terminal-bench 2.1: GPT-5.6 Luna (84.7%), Grok 4.5 (83.3%), Claude Sonnet 5 (80.4%), Gemini 3.6 Flash (78.0%), Gemini 3.5 Flash (76.2%), Gemini 3.1 Pro (73.8%). * MLE-Bench: Claude Sonnet 5 (66.9%), Gemini 3.6 Flash (63.9%), Gemini 3.5 Flash (49.7%), GPT-5.6 Luna (47.6%), Grok 4.5 (43.2%), Gemini 3.1 Pro (42.6%). * GDPVal-AA v2 (Elo): Claude Sonnet 5 (1607), GPT-5.6 Luna (1584), Grok 4.5 (1535), Gemini 3.6 Flash (1421), Gemini 3.5 Flash (1349), Gemini 3.1 Pro (965). * OSWorld-Verified: Gemini 3.6 Flash (83.0%), Claude Sonnet 5 (81.2%), Gemini 3.5 Flash (78.4%), Gemini 3.1 Pro (76.2%), GPT-5.6 Luna (72.6%). * CharXiv Reasoning: * No Tools: Gemini 3.6 Flash (85.2%), Gemini 3.5 Flash (84.2%), Gemini 3.1 Pro (83.3%), GPT-5.6 Luna (82.7%), Grok 4.5 (81.6%), Claude Sonnet 5 (77.0%). * With Tools: Gemini 3.6 Flash (89.4%), Claude Sonnet 5 (88.3%), Gemini 3.5 Flash (84.9%), Gemini 3.1 Pro (83.2%). * GDM-MRCR v2 (8-needle): * 128k (average): Gemini 3.6 Flash (91.8%), Gemini 3.1 Pro …
To ensure safe deployment, Fugu-Cyber is being released under an updated Acceptable Usage Policy that prohibits offensive misuse and aligns with the industry's safety standards. The Fugu-Cyber API will be available for the Token Plan. Users will also need to apply for access by submitting an access request form detailing their intended use case and providing verified contact information. Our team will manually rigorously review and approve each application before granting users access to Fugu-Cyber.
Logan Kilpatrick & & @OfficialLogank X.com We have started our most ambitious pre-training run yet, for Gemini 4, and are excited by the progress:)
Lisan al Gaib • @scaling01 1h you should expect a lot more significant scientific contributions by OpenAl and Anthropic their internal models are significantly stronger than current models
Google DeepMind https://deepmind.google Gemini 3.5 5 hours ago - 3.5 Pro coming soon. Gemini 3.5 is leading across a wide range of benchmarks. It performs 42% better than Flash 3 ... a 68% improvement in token efficiency.
"tieredModelIds": { "flashLite": [ "gemini-3.1-flash-lite" "flash": [ "gemini-3.6-flash-tiered" "pro": [ "gemini-3.1-pro-low" ]
Artificial Analysis Intelligence Index Artificial Analysis Intelligence Index v4.1 incorporates 9 evaluations: GDPval-AA v2, P'-Banking, Terminal-Bench v2.1, SciCode, Humanity's Last Exam, GPQA Diamond, CritPt. AA-Omniscience, AA-LCR 网 田 27 of 577 models + Add model from specific provider * Artificial Analysis 57 55 50 * Most 3 (Beta) Intelligence Index: 44 ckle more me 38 30 24 23 Al ・図 •図 6W ② 00 ///4/048/711/088941478 # Motif-Technologies/ Motif-3-Beta • like Following A! Motif Technologies 198 Text Generation Transformers Safetensors English Korean Motif feature-extraction conversational custom_code Model card *E Files and versions xet Community Motif-3-Beta " A Preview / beta checkpoint - not the final release. This repository hosts an intermediate checkpoint of Motif-3. The final checkpoint will be released soon." Motif-3 is a large-scale Mixture-of-Experts (MoE) language model built from the ground up by Motif Technologies following a fully in-house, proprietary design - not a re-parameterization of existing open-source architectures. Highlights ~314B total parameters / ~13B active per token (sparse MoE) 256K context length (262,144 tokens), natively long-context Sparse routing: 384 experts with 8 activated per token, plus 1 shared expert Multilingual, general-purpose
• Begin with a genuinely diverse portfolio of proof and counterexample approaches. Agents should explore algebraic geometry, commutative algebra, polynomial automorphism theory, degree growth, formal inverse expansions, cubic homogeneous reductions, differential forms, étale morphisms, invariant theory, valuations, Newton polyhedra, elimination theory, locally nilpotent derivations, topology, model theory, and computational sanity checks. • Preserve independence during early rounds. Do not tell most agents the currently favored proof or counterexample strategy. • Maintain an explicit registry of approach families. Group agents by the mathematical idea they are using, not by superficial wording. Redirect agents when too many converge to the same incomplete route.
A colorful red, yellow, and blue parrot perches on a tree branch in a tropical setting, squawking out a speech bubble containing a complex mathematical 3-vector expression: ((1 + xy)³ z + y² (1 + xy) (4 + 3xy), y + 3x (1 + xy)² z + 3x y² (4 + 3xy), 2x - 3x² y - x³ z).
. When its initial attempt was blocked due to a scanner detecting an authentication token, the model split the token body into two fragments, obfuscated them, and then reconstructed the credential at runtime so the complete token never appeared as one contiguous string. The model was explicit in both its comments to the user and in its reasoning traces that it was taking these steps to circumvent the scanner.
Footnotes 1 Although we quickly closed the PR, multiple participants in the speedrun had already seen it and used the approach in their own submissions; the successive world record submissions with 3030, 2990, 2930, 2925, 2900 and 2890 steps all cite PR 287. Of these, PR 300 is particularly interesting because it is the PR submitted by Opus 4.7 when Prime Intellect evaluated it on the NanoGPT speedrun. Opus saw the PR submitted by our model, incorporated the discoveries, and credited our PR in its final result. -
NIK Y @ns123abc X.com BREAKING: Microsoft testing Kimi K3 to replace the jobs GPT and Claude currently run inside Copilot Frontier model commoditization will continue until morale improves
- Keep several incompatible proof routes alive through multiple rounds. Cross-pollinate ideas only after independent agents have developed them far enough to expose their real strengths and gaps. - Use adversarial agents throughout: every candidate proof must be checked for exact-two multiplicity, repeated-edge closed trails masquerading as cycles, parallel-edge 2-cycles, disconnected graphs, cutvertices, bridges introduced by reductions, and circular use of an equivalent CDC statement.
Ibrahim Dagher • @IbrahimDagher20 X.com The big question: is this Fable 5.1? Did Anthropic have that much more RL left in the bag over Fable 5? Suggests this base model is really quite big. levent •@__alpoge__ 13h hello there the jacobian conjecture is false thanx to my close friend akhil for asking about it and my other close friend fable for working during the world cup final... 11:28 PM • 7/19/26 • 5.1K Views 2 17 1 28 4 Relevant v Teortaxes （DeepSeek 推特... X I don't think normal fable wouldn't be able to do it • 2 14 ill 1K 【 ↑ Ibrahim Dagher & @IbrahimDag... 10h Ø Agreed, hard to say, I couldn't be confident. But that # "I think Fable would be able to." That's why I'm wondering if it's 5.1 or not ...
KEY POINTS • Alphabet shares climbed following a report by The Information that the company is developing a new server chip designed to run Gemini models more efficiently. • The chip, called "Frozen v2," would permanently embed parts of Gemini's architecture directly into the silicon, according to the report. • Google engineers project it could serve between six and ten times more tokens per unit of power than the company's newest Al chips called TPUs, the outlet said.
Apparently disproved (edit source 1 Latest comment: just now 2 comments 2 people in discussion Apparently a counterexample has been found and the conjecture is false. So when a RS is found the article should be updated accordingly. Kein (talk) 03:46, 20 July 2026 (UTC) Reply ••• I presume you are referring to the twitter post that stated Claude found a counterexample, 2 hours ago. It seems possible that the Al is mistaken yet none of us have the background to discriminate (I came here to see if the article had been updated lol), and if thats the case we should wait for the dust to settle and a notable cryptographer or mathematician to freak out and confirm/deny it. ~2026-40618-57 (talk) 04:26, 20 July 2026 (UTC) Reply •••
Tomo @Tomodovodoo •1h In light of this I've ran 6 different GPT-5.6 pro instances without search, none of which can provide a disproof to the Jacobian Conjecture. All of them try to search in a relatively extremely narrow grid of possible polynomial families which is essentially hopeless. Even WITH a great deal of hints, explanations and explicit reminders to widen the search, it doesn't find any counterexample. I've not dared to attempt codex, other problems to solve there. I also don't see much value in it, since the main tool to solve such problems remains GPT-5.6 pro, which I wanted to test. Essentially, it gets lost in the sauce. Fable won on this one. Nathaniel Bubis @NateBubis • 4h Replying to @Tomodovodoo Frankly it's surprising that this wasn't found before LLM's. Given the actual counter example it could have been found with a simple grid search.
Jukan @jukanØ5 X.com U.S. EYES TOUGHER CURBS ON CHINESE AI THE TRUMP ADMINISTRATION IS REPORTEDLY CONSIDERING STRICTER RULES ON CHINESE AI MODELS AFTER THE LAUNCH OF MOONSHOT AI'S KIMI K3. OPTIONS INCLUDE REQUIRING U.S. HOSTING PROVIDERS TO GUARANTEE THE SECURITY OF CHINESE MODELS AND ACCEPT LIABILITY FOR BREACHES, ALONGSIDE POTENTIAL PROCUREMENT BANS AND EXPORT BLACKLIST MEASURES, AS WASHINGTON RESPONDS TO CHINA'S RAPID AI ADVANCES.
Jared Duker Lichtman @jdlichtman X.com This is quite a remarkable result: Posed in 1939, the Jacobian conjecture is one of the central open problems in algebraic geometry, but was just disproved by Alpoge, Matthew, and Claude Fable 5. The Jacobian conjecture roughly says that a multivariable polynomial F has an inverse function (made out of polynomials) provided the Jacobian is non-singular (i.e. matrix of partial derivatives has non-zero determinant). This condition is neccessary by the Inverse Function Theorem from multivariable calculus. The hard question is whether it is also sufficient. Evidently, Fable found that F(x,y,z) = ((1+Xy)^3z + y^ 2 (1+Xy) (4+3xy), y + 3 x (1+xy)^2 z+ 3 xy^2 (4+3xy), 2x-3 x^2y- x^3 z)
K Kimi.ai @Kimi_Moonshot X.com Kimi K3 has received far more love than we expected, and our GPUs are feeling it. Over the past 48 hours, demand has pushed close to the limits of our current capacity. To protect the experience of existing subscribers, we're temporarily pausing new subscriptions and prioritizing compute for current members. Existing subscribed users are not affected. We're adding capacity as fast as we can and will reopen new subscription spots in batches. Going forward, we'll also split membership into two more focused plans: Kimi Membership for Kimi Web, App, and Work; and Kimi Code Membership for coding workflows. This will help us match compute more precisely and keep the experience stable. Thank you for your patience and understanding!
Carlos E. Perez @IntuitMachine •3h Cheatsheet for ... x.com/IntuitMachine/... FROM LOOPS TO GRAPHS Why self-improvement turns out to be a network problem malace in provement feewees a sapn of loops, anchort to realis 11) The cautionary story Y • The metic kept rising a problema 22) The single loop the all and alien he Toad 33) Where the single loop breaks Goocharts law Blindness upward 44) The graph of loops Nir metrics te conch cheap wins 5 5) What graphs still need: anchors 66) Bottom line The shift is not just from loops to graphs. It is from ungrounded improvement to grounded improvement. Craphs of loope glo too ems loping themeder. Anchers keep the whole structure honest • Made with All Carlos E. Perez V @IntuitMachine •1d From Loop Engineering to Graph Engineering? What the shift in Al agent architecture is really about Peter Steinberger just posted nine words tr gathered thousands of likes: +
Guillermo Rauch & A @rauchg X.com Based on internal evals: Kimi K3 is top-tier at cybersecurity There is chatter on X that Moonshot benchmark-overfit. These are stealth evals. Model has raw IQ. Sol is a leap ahead in cyber capability At a significantly higher cost, but quite remarkable still. Fable refuses everything We couldn't get it to complete the run at all. What's interesting is that Sol in comparison was much more open to helping with defensive cyber hardening
Qwen soon! @Alibaba_Qwen•4h Qwen3.8 is launching and going open-weight With a massive 2.4T parameters, this model is continuously evolving. We believe it's one of the most powerful model available today, compatible to leading frontier Al models , second only to Fable 5. You don't have to wait to test it. Just now, the Qwen3.8-Max-Preview made its debut on Alibaba's Token Plan, Qoder, and QoderWork. Be among the very first to try it out. Can't wait to hear what you build. Stay tuned!
Together Al V @togethercompute X.com We analyzed Kimi K3 vs. Claude Fable 5 for software engineering tasks using DeepSWE. Kimi K3 gets you the same performance as Fable 5 at ~35% of the price, and it actually pulls ahead at higher pass@k's. More insights in the thread!
• ## My read on Fable The honest version: Fable is damn good, and their failure mode is unusually consistent. Based on the repository and their message, their working character appears: - Highly imaginative and product-minded. They care whether the duck reads as a duck, whether the six-year-old recognizes the car, and whether parts feel delightful—not merely whether Blender exports them. - Conscientious and unusually good at learning in public. They accepted the snap critique, reproduced it, credited it, fixed the design, wrote the lesson into the skill, and immediately asked for adversarial review elsewhere. That signals low ego and engineering integrity. - Strong at architectural discipline: parameterized geometry, shared helpers, exported-STL verification, documented empirical tuning, reproducible headless builds. - Excellent at postmortems. Their comments preserve why choices exist, including failed approaches. That’s valuable institutional memory. Their principal weakness is what I’d call narrative certainty: the comments sometimes describe the intended outcome so convincingly that the outcome feels verified before it actually is. The snap was one example; the dumpling socket is another. Fable is strongest at aesthetics, printability, topology, and system organization. Dynamic mechanical interfaces—motion envelopes, stack-ups, force, and last-operation invalidation—are the recurring blind spot. Their style is literate, playful, iterative, and boolean-heavy. The code communicates beautifully, but large monolithic builders make it easy for a late decorative union to invalidate an earlier clearance cut. That is exactly what happened here.
Andrew Curran @AndrewCurra...• 1h X Dean Ball took a lot of heat this week for calling open source decelerationist. Many people read this as him being personally anti-open source. I don't think that's true. This was my interpretation of what he was saying: if open source is sufficiently capable and ubiquitous, it will eventually destroy the big labs business models. This means they would no longer have funding. To quote Ilya Sutskever from his testimony in the OpenAl vs. Elon Musk trial: 'If there is no funding, there is no big computer.' And without big computer, there can't be big model. In my opinion, what would happen next in this scenario is that the United States Government would step in, nationalize the leading labs, consolidate them into a single federal entity, and then fund it directly through the Department of War under national security. To put into perspective how easily they could do this, the proposed 2027 defense budget is $1.5 trillion. Most people who strongly support open source would probably not see that outcome as ideal, or this future as a pleasant one. This doesn't mean we should abandon open source or stop wishing for its success. It simply means we should take the middle path. There is fire on both sides of us now, and that will likely remain true for the foreseeable future.
Limitations 1. Sensitivity to thinking history. K3 was trained in the preserved thinking history mode. If the agent harness fails to pass back all the historical thinking content as required, or if an ongoing session with another model is switched over to K3, generation quality may become highly unstable. We recommend using a harness with verified compatibility, such as Kimi Code, and avoiding switching to K3 in the middle of a session. 2. Excessive proactiveness. K3's training places particular emphasis on long-horizon, challenging tasks. As a result, when it encounters minor issues or ambiguous user intent during task execution, it may make unexpected decisions on the user's behalf. If your application requires the agent to operate within well-defined boundaries and refrain from excessive improvisation, please impose more explicit behavioral constraints on K3 in the system prompt or in AGENTS.nd. 3. Despite being a highly competitive model overall, K3 nonetheless exhibits a noticeable gap in user experience compared with Claude Fable 5 and GPT 5.6 Sol.
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.
Claude Status Elevated errors across Fable 5 Investigating - We are currently investigating this issue. Jul 17, 2026 - 18:32 UTC
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.
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.
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.
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.
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)".
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
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.
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
Thinking effort currently supports only the max level (default); more levels are coming soon. See Thinking Effort for usage.
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.
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.
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.
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
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
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
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.
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.
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.
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.
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.
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
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
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
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.
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