Ntovas G

This chart I believe is far more important for the future of AI

@ariaurelium I use 512k when I mean 524,288

people sometimes use "64k" to mean 65,536 because "65k" sounds weird after "32k"≈32,768 at what power of two can you no longer do this. like would people get mad if I said "512k" and meant 524,288

@osanseviero Usually the average model performs great if you make the problem in distribution. The benchmarks measure the performance of the model in a problem agnostic harness that is not the case in small open weight models.

2 years later, still relevant 👇

@MiaAI_lab I believe the "new gemma" is not out yet, the chat template change is mentioned as just one of the changes.

The “new” Gemma 4 update is just a chat template change — the model is the same. But it does fix some real issues that were affecting tool calling and reasoning in multi-turn scenarios. The most meaningful addition is the "preserve_thinking" flag (similar to what Qwen has), which should help preserve reasoning across tool calls and improve reliability in agent workflows. However, when it comes to actual performance, Gemma 4 31B still falls short of Qwen 3.6 27B in agentic/tool-use benchmarks. It’s also slower. At this point, unless you have a strong preference for Google, there isn’t much reason to choose Gemma 4 over Qwen 3.6 for agentic work. Here is the Tool Eval bench comparing the two 👇

@cheatyyyy So, we are going faster than the AI 2027 predictions. Nice

@vineholdersuper @bryan_johnson Sorry if thay is not clear, but I am not interested in relationships with men, so even if you found me pretty, it wouldn't mattered to me. Thanks.

@ntovas_gi @bryan_johnson Spiritually people always have ugly physiognomy

The world wants me to die. My incurable disease diagnosis became global news. It was omnipresent on social media and 1,900 articles were written in a matter of days. Many were saddened. However, joy dominated the commentary. People pointed to schadenfreude, the pleasure of another's failure. Yes, there’s that. There is a special place in people’s hearts that loves to see others fail, especially when that person’s presence threatens their own psychological stability in some way or helps them feel better about themselves. But, if you look over the social media commentary about me, you’ll see that pattern: “he deserved it.” I deserved it because I challenged death. The crowd was running a deeply rooted psychological script that represents the oldest, most deeply embedded stories of human culture. This was the first story ever written down, 4,000 years ago. Gilgamesh sought eternal life after losing someone he loved, only to have the plant of youth stolen by a serpent as he bathed. Leaving him to accept his mortality. Asclepius became so skilled at rejuvenation that he raised the dead. As punishment, Zeus struck him down with a thunderbolt to enforce life and death authority. This is the story of Jesus. Pontius Pilate offered a choice between a thief and the immortalist, and the crowd demanded the execution. People need this story conclusion to keep themselves sane. The challenger must lose and the loss must appear deserved. It’s a shield of self preservation. For if death is inevitable, their existence and that of their loved ones is justified and unavoidable. If death is not inevitable, nothing about their reality is safe. I occupy the same philosophical and archetypal position as Gilgamesh, Asclepius and Jesus. This statement will draw outrage and accusations of blasphemy, hubris and narcissism. Nevertheless, it’s the pattern that has repeated itself for thousands of years. Death has been the omnipresent concern of the human race. It encapsulates our greatest fears, joy and curiosities. The discourse around it changes over time; however, the fundamentals remain unchanged. What’s different about this moment, that is unlike any other moment, is that physical death may no longer be inevitable. What if I didn’t deserve it? And what if I am your ally, and not a threat?

@mikewadhera No, you don't understand. It's a static site generator with an AI assistant!!! We are talking about technological breakthroughs here.

bro you guys run a static site generator

@oleavr It's time for Frida+AI 🤣 btw kudos for possibly the best tool in existence today ❤️

How time flies ☺️

@OGNodeFather @__tinygrad__ @scaling01 Trash compared to what?

@OGNodeFather @__tinygrad__ @scaling01 Any model released in 2026 bigger than 4b, would be considered AGI 6 years ago.

lol there's no window. there is a bubble, but there's no window. you would have been posting about how the Apple II (1977) is the goat but with enough resources IBM could catch up with the IBM PC (1981). The Commodore 64 dropped in 1982. The NES dropped in 1983. All those computers have less power than the touch panel controller in my cell phone. As is was for computers, as it will be for LLMs. "Mythos class" will run on a midrange laptop in 6 years. training it will be in the realm of hobbyists. This is going to be the most proliferated technology since the tamagotchi.

@zekramu The secret is that it doesn't matter if it is slop. I have worked on a company where some parts of the business was running on .vbs scripts, last changed 15 years ago. Nobody cares about your perfect monads out of your circle jerk.

here’s a little known secret: software was always slop, long before llms. software engineers are way to fucking prideful of the text they put on a screen to admit this.

@AnthropicAI @AnthropicAI can you fix claude code? Today is like talking to a 1bit quant of Opus, speaking Chinese, forgets how to wait for sub agents, etc.

Claude Fable 5 will be available again globally tomorrow. After a series of productive conversations with the US government, we're redeploying the model with a new set of classifiers to target and block more cybersecurity tasks. In the near term, some routine tasks like coding and debugging will fall back to Opus 4.8. We’ll continue to refine these classifiers over the coming weeks to reduce false positives and better distinguish genuine misuse from legitimate requests. We’ve also begun drafting a consensus framework—with Amazon, Microsoft, Google, and other Glasswing partners—for assessing the severity of AI jailbreaks and how AI developers should respond to them. We invite other industry partners and model providers to join us in this effort. Finally, we’re scaling up our collaboration with the US government on model testing and safeguards. This will include pre-release access to models and safeguards for evaluation, information sharing on jailbreaks and misuse, and dedicated resources for joint research. Thank you to our users for your patience, and to our partners across the government, industry, and the research community who worked alongside us to make Fable 5 available again. Read our full blog: anthropic.com/news/redeployi…

@hsu_steve Opus 4.8 and the disabled Fable are so far ahead, that companies are willing to pay 10s of thousands of dollars per employee to use them. I love open models and they are great for small tasks, but there are not even close to 15+ api hour horizon of Opus.

Interesting GLM rumor from China sources: "Apparently Z.ai has an internal router behind their coding plan that routes your Claude Code query to GLM or Claude depending on if the classifier thinks it is in distribution or not. If it is OOD and also high value, it will route to Claude and then add the trace to the distillation dataset." Seems hard to defeat this since these are real user queries and not contrived. The accounts that generate this distillation data can be made to look like ordinary user accounts but with higher concentration of OOD queries. This looks like a fast-follower strategy that will keep a weaker lab in the game at a lower price point per unit of intelligence or per token.

@staysaasy Lol, Salesforce is already a bloated platform with most things partially working and most companies pay an army of consultants to feed them excuses every week. You can write the crm your company NEEDS in a couple months if you are semi-competent.

“I can rebuild salesforce in a day” First - no you can’t Second - when you’re done building the 1% of functionality you call done, now your entire GTM team needs to onboard into it. Who is gonna do that, you? You immediately have a full time job of enabling the team on your inhouse CRM and your salary disappears. Now you hire a new EMEA head of sales and he says he can’t hit his target without five features Salesforce built a decade ago. You vibe code them 20% of the way and ship them too late. They miss your quota and Blame you. The CEO starts having panic attacks. Now you have an incident at 2am. Your entire APAC sales team can’t work for 8 hours while you sleep off your late night goon sesh. Immediately the company requires oncall for This project. You hire 3 more engineers with a total carrying cost of 1.2M, which is double what you were paying Salesforce. Your CTO starts having heart palpitations. Sales misses another quarter and blames your CRM again. Meanwhile other people in the business are vibe coding CRMs to compete with yours. Alex’s sucks they say, I can vibe code better a day. Now 10% of company resources are spent demoing CRM tools. You exercised a bunch of options earlier in the year and have a huge AMT bill. The company promised a secondary at next raise but they just missed another quarter and will not do a secondary. The company decided to scrap your CRM and go back to Salesforce. They thought it would fix everything but it didn’t, because the real problem was that the company was so out of ideas for revenue growth that they let you vibe code a CRM as a seemingly good idea. Another quarter missed. Layoffs ensue. You can’t pay your AMT bill and the IRS foreclosed on your house. The business gets acquired underwater and you get nothing. You’re now living under a bridge in Philly wearing nothing but an Eagles starter jacket, porky pigging around and selling your body for root beer and tranq. Some guy drives by in a corvette and splashes a puddle on your bare ass. You look over the license plate says BUYSAAS. You fade into unconsciousness thinking that you should have just focused on your regular OKRs instead of constantly trying to side project some bullshit.

@an_engineer_log How is Kia still relevant in a world where Ferrari exists?

How is Gemma 4 still relevant in a world where GLM 5.2 exists ? x.com/OfficialLoganK…

@5_utr Everything is a sigmoid, if you force it to be a sigmoid. On the real world we are in an exponential yet.

AI scaling is not exponential. It never was. I fit a logistic model to general capability vs compute across every major model generation — AlexNet to GPT-4. R² = 0.98. The asymptote is not infinity. Never was or will be. Thread 🧵

@Espen_Antonsen @deedydas I didn't said anything about replacement. I said that having code that is shit and nobody wants to touch was the norm many years before the LLMs.

@ntovas_gi @deedydas So a service that runs for 10 years without problems should be replaced with something that needs to be patched every 10 hours? That’s an improvement?

Most software engineers are facing an identity crisis bordering on depression. As CTOs aggressively evangelize tokenmaxxing, a class divide ensues. The lazy. The lazy push code. They don't write it. They don't manually test it. They don't even read it. They're on autopilot. See Jira ticket, prompt for task, submit code. Many of them are barely on their computer the whole day. A comment on the PR asking why they did this? The lazy ask AI. A Slack message? The lazy ask AI. Need to prepare for standup? The lazy ask AI. As long as it sounds enough like them and isn't detected. Some of the lazy are even overemployed, and work multiple jobs. The lazy smart ones get away with this, and even rewarded. After all, software engineering for the lazy is just a dance to convince your colleagues you're smart and hard working. The craftsmen. The craftsmen are tired. Very tired. 15 PRs in queue. Slack blowing up. The entire burden of review falls on the craftsman. The burden of understanding. They try. They work their way through the code, thoughtfully commenting to improve what ships. The response? A lazy: "That's a clever idea! You're absolutely right." with an incorrect change. It's fine, the craftsman says. I can fix them. They write a doc urging his colleagues to be better. The next day? 20,000 line PR to review. Day after day, their workload grows. Bugs seep into production. No one seems to care. Another round of AI is thrown at it. Their animosity to their colleagues rises. Eventually, they give up. It's just not what it used to be. The craft they loved is dead. They eventually wake up, a lazy. This isn't all companies. Many companies are genuinely more productive, adopt the right set of principles and practices around AI development and have highly talented teams that trust each other. It tends to happen in bigger companies that are 10+yrs old with a higher talent variance. But it happens. A lot.

@tts_ff2 @deedydas That was the case always, with worst code, no tests and all the human stupidity in the code. At least now you have something to work with.