Sabitlenmiş Tweet
Neto
2K posts
Neto retweetledi
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
TL;DR: frontier, open-weight cybersecurity capability is here. Try it on deepsec.sh for defensive purposes.
Malte Ubl@cramforce
We ran Kimi K3 on a private cybersecurity benchmark. TL;DR: Kimi K3 is the workhorse for cyber security tasks at great recall/precision/price. GPT 5.6 is best recall/precision but at 7x higher cost per run. For context, Deepsec.sh is an open-source cyber harness designed for finding vulnerabilities in large codebases. The eval runs deepsec on an undisclosed open-core application at a git sha before a large number of security issues were fixed. This is a secret eval that cannot be directly benchmark-maxxed. S-Tier: GPT 5.6 Sol: By far the most thorough analysis, but coming in at over 7x the price of the runner up. Best price/recall: Kimi K3. Next tier of recall at a good price Best price at good recall: GLM 5.2 (40% lower price than Kimi K3) GPT 5.5: Only recommended with subscription or high-discount API price. Similar recall to Kimi at much higher list price. Opus 4.8: Only recommended with subscription or high-discount API price. Similar recall to GLM 5.2 at much higher list price. Fable 5: 100% refusal rate. Cannot be used for security analysis. Sol on a large code base will quickly get into 6-figure pricing. This is still affordable relative to the risk of letting security issues unfixed or paying bug bounties. I'd recommend using Sol for a one-time baseline and then using Kimi K3 for continuous analysis. When using open-weight models, make sure to use an inference vendor that supports zero data retention.
English
@deanwball it’s one of the most annoying things about social media. general absence of good will that becomes a vicious cycle: most major accounts talking their book, presuming that means everyone is.
English
I’m afraid to tell you that it is effectively impossible to do the kind of writing I used to do on this website, not because anyone at OpenAI censors me but because of the sheer volume of hostility I get for sharing my analysis as a frontier lab employee.
I enjoyed writing quick takes on this website for one basic reason: I could get rapid feedback on my own ideation process in real time. Post the early version of the take here, see the criticism; then refine, sharpen, and repeat. Unfortunately now that feature of this site is gone, because the feedback I get is now almost exclusively colored by resentment at the fact that I work at a frontier lab or other forms of hatred for my employer. The feedback signal is essentially useless now, so writing on here is not fruitful for me anymore.
Literally everything I write now is responded to with “of course you said that because .” I am truly just writing what I think and would have written anyway, but everyone reads what I say in the shrieking tone of “this is what openai thinks!!!!” (to be clear, my posts are not what openai thinks). This is an unpleasant and more importantly unproductive pattern for me.
I anticipate that the shape of this account will change significantly as a result. I do not currently know how. It will not become a LinkedIn feed. It will change in some other way. It will no longer be a real-time accounting of my own thinking as it develops, since this is precisely the thing that seems impossible to do now. That will have to shift to private channels.
English
Neto retweetledi
Introducing Honen
Turn your company context into interactive courses that self-improve over time
We partnered with NVIDIA to bring AI literacy to 250,000 learners
We're ensuring no person or company gets left behind
Get started at honen.com/contact
English
Neto retweetledi
OH: “i’ve switched to Kimi from claude for a bunch of work. it’s just so much more fun because it just does the thing instead of lecturing you”
Woke lobotomized models are the enemy of American competitiveness.
English
Eu particularmente acho que o subsidio da China é do governo (para treinar a LLM) e nos EUA é o dinheiro privado. Mas os 2 são subsidios mas não vejo isso como um problema. O preço da venda do token só existe pq alguém subsidiou a fase de aprendizado da LLM, seja governo ou privado.
Português
Vou pinçar essa resposta para expandir mais sobre a questão dos subsídios na IA já que há muita confusão com o termo.
Por definição, todo P&D é subsidiado, uma vez que o produto/serviço ainda não existe. Esse "subsídio" pode vir de diversas fontes:
- Programas governamentais
- Incentivo fiscal
- Receita de outros setores da empresa
- Caixa levantado por investidores
Uns são mais "sustentáveis" que outros. Por exemplo, é muito mais estável subsidiar um produto com receita interna do que depender de uma canetada do governo, ou de dinheiro barato de VC.
Mas se você aplicar esse escrutínio com rigor, nada vai restar na fronteira da tecnologia. Tudo é de certa forma financiado por outra atividade.
Quando digo que a IA não se sustenta por subsídios, não me refiro aos novos modelos que virão. Esses provavelmente serão sempre subsidiados – de novo, como qualquer pesquisa.
Me refiro explicitamente ao custo de inferência, ou seja, se é economicamente viável rodar o modelo no meu dia a dia, consumi-lo como algo que não preciso me preocupar se amanhã ele ainda vai existir.
Isso já não era verdade há tempos, com a diferença de que o gap dos modelos open source tops de linha para os modelos "privados" (e portanto, esses sim podendo ser desligados a qualquer momento, vide Fable vs governo americano) era de ~6 meses e com o K3 caiu para ~2 semanas.
Ou seja: se amanhã o Trump proibir o Fable e o GPT 5.6 novamente, e o Xi Jinping decidir nunca mais colocar 1 yuan em IA, já existe um modelo que há 2 semanas era considerado top 2, economicamente viável, e por cerca de metade do preço do que era praticado.
Neto@edmcneto
@rafaquint Mas Rafael, existe um certo subsidio da China, concorda?
Português
Neto retweetledi
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 AI 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 AI -- while addressing risks in a targeted way -- or we’ll watch our lead evaporate.
Arena.ai@arena
Big news: Kimi-K3 by @Kimi_Moonshot is now #1 in the Frontend Code Arena with 1679 pts, surpassing Claude Fable 5. This is a 17-place jump from Kimi-k2.6 (#18 -> #1). In Frontend, Kimi-K3 ranked #1 in 6 of 7 domains: Brand & Marketing, Reference-Based Design, Data & Analytics, Consumer Product, Simulations, and Content Creation Tools, landing #2 only in Gaming behind Fable 5. The full model weights will be released by July 27. Congrats to the @Kimi_Moonshot team on this major milestone!
English
A melhor notícia não é o Kimi K3, o preço, ou o benchmark – mas sim que a partir de hoje um modelo da capacidade do Opus 4.8, que há 2 semanas era top 2, já é economicamente viável por ~metade do preço. E ainda tem gente que acha que a IA só se sustenta por subsídio...
Artificial Analysis@ArtificialAnlys
Kimi K3 scores 57 on the Artificial Analysis Intelligence Index. Its intelligence is comparable to Opus 4.8 and GPT-5.5 but remains behind Fable 5 and GPT-5.6 Sol. Moonshot AI has expressed plans to release the 2.8T parameter model's weights, which would make it the leading open weights model Key results: ➤ Strong agentic task performance: @Kimi_Moonshot's Kimi K3 reaches an Elo rating of 1668 on GDPval v2. This is a marked improvement over K2.6’s 1190, surpassing GLM-5.2 (1514), GPT-5.5 (1494), and Claude Opus 4.8 (1600). However, it still lags behind Claude Fable 5 (1760). Kimi K3 also scores an impressive 53% and takes the #1 position on AutomationBench-AA, our implementation of Zapier’s Agentic SaaS workflow evaluation. ➤ Second-highest performance on AA-Briefcase (agentic knowledge work): On our private long-horizon knowledge work evaluation, Kimi K3 reaches an overall Elo of 1547, +732 points from Kimi K2.6 and behind only Claude Fable 5. It is well-rounded: its rubric scoring and analytical quality almost reach Claude Fable 5’s scores, while GPT-5.6 Sol continues to outperform other leading models on presentation quality. ➤ Set to lead open weights models once weights are released: Moonshot AI has not yet released the weights but expressed plans to do so. Once available, Kimi K3 would clearly lead other open weights models including GLM-5.2 (51) and DeepSeek v4 Pro (44). However, at 2.8T parameters, it is significantly larger than its open weights peers (eg. GLM-5.2 at 753B params and DeepSeek V4 Pro at 1.6T), as well as the Kimi K2 to K2.6 models (1T params). ➤ Cost per task ($0.94) is similar to GPT-5.6 Sol ($1.04), ~1/2 the price of Opus 4.8 ($1.80) and higher than open weights peers: Moonshot AI’s pricing for K3 is significantly higher than their K2 pricing (K3’s output token price is $15/1M tokens while K2.6 was $4). This positions the model as cheaper on a cost per task basis than Opus 4.8, similar to GPT-5.6 Sol ($1.04) and more expensive than open weights peers, GLM-5.2 ($0.32) and DeepSeek V4 Pro ($0.04) ➤ Improved token efficiency alongside higher intelligence: Kimi K3’s token usage on the Artificial Analysis Intelligence Index decreased significantly, using 21% fewer output tokens than K2.6. The new model used approximately 132M output tokens to complete all nine evaluations, compared to approximately 166M for K2.6, while achieving higher scores. ➤ Native multimodal capabilities: Kimi K3, like K2.6, is released with native image and text multimodal input. If weights are released, this will position Kimi K3 as one of the leading open weights models with multimodal input capabilities Other model details: Context window: 1M Size: 2.8T total parameters Pricing: The first-party API is priced at $3.00/$15.00 per 1M input/output tokens, with cached input discounted 90% to $0.30 per 1M tokens. Modality: Native multimodal input supports text and images, and the model remains text-only for output. Accessibility: Accessible at launch through Moonshot’s first party API. Model weights are not yet released but Moonshot AI has expressed plans to do so.
Português
Neto retweetledi
Today we share the worldview behind our mission.
Human values don't average out. Local knowledge can't be centralized. The good future has many AIs, raised in different places, shaped by the people they serve, disagreeing with each other the way we do.
thinkingmachines.ai/blog/the-futur…
English
Germany has launched one of the world's best open-source AI models.
Soofi S, made by the Soofi consortium, is a 30B parameter model fully trained in Europe and tops the ranking for open-source AI.
Huge moment for Europe, and finally some competition for Chinese open-source AI.
English
Neto retweetledi
Cursor’s founder on why AI makes the next engineer more valuable, not less.
The New York Times spends ~$150M-$200M a year on software R&D, even though most people do not think of it as a software company.
Professional software engineering is still far from solved because building a new codebase is not the same as safely migrating an existing system like Rippling’s 30M-line codebase, with years of logic, dependencies, and customer workflows behind it.
Cursor’s 20-person support team handling millions of daily users clearly shows that AI leverage for small teams is real.
But the scope of software work is still vast, so the ROI of adding the next strong engineer is higher, not lower.
English
O que tem de gente sendo levada da Vercel pro Cursor não é brincadeira. RH do Cursor tá a base da baciada na Vercel. Bizarro.
Português
Cafe Cursor is coming up in these cities:
• LA - 7/15
• Philly - 7/16
• Lisbon - 7/17
• Zambia - 7/18
• Salvador - 7/20
• SF - 7/20
• Medellin - 7/22
• Rosario - 7/24
• Dallas - 7/26
• Cologne - 7/30
• Bali - 8/3
• Bangkok - 8/9
• Casablanca - 8/9
More being added
English
@glaucia_lemos86 Oi @glaucia_lemos86 tenho te acompanhado bastante e estou impressionado com a solidez de seu conhecimento. Parabéns.
Português
Neto retweetledi
I built an LLM from scratch. Here's everything I learned along the way.
English
Neto retweetledi
My skills repo has 160K stars, 7.5m downloads...
...and no tutorial.
So, here it is. Watch me walk through the essential skills:
- /grill-with-docs
- /to-spec
- /to-tickets
- /implement
- /code-review
It's the whole flow, end-to-end. Enjoy:
English
Neto retweetledi
Neto retweetledi
We've usually stayed away from model comparisons but 5.6 vs Fable is a unique situation
We've never had a case where the team is so completely convinced on which one is better
Here's the timeline of our experience with it
- We test early versions of 5.6 for a couple of weeks and have a great time, it feels like a step change improvement, enabling new workflows
- We get to try Fable and don't think it's not as good, I personally would take this experience with a grain of salt, there tends to be a bias when trying a new model when you already like another
- Fable and 5.6 are taken away because of the regulatory issues
- Our team is literally depressed that 5.6 is gone, we are looking for anything that could even partly replace it
- Fable comes back, and here's where it gets interesting, you would think Fable would be enough, but no, the team is still depressed that 5.6 isn't available
- Then 5.6 comes back and it's immediately clear that it's just way better than Fable
This situation was unique in that it was the closest we've ever gotten to having an unbiased comparison of two models
English
Neto retweetledi
Four of AI's most confusing terms explained:
An agent is just a model, harnessed, in an environment.
English