Live data from Hacker News

The path to ubiquitous AI (17k tokens/sec)

taalas.com

201–210 of 471 posts

Re: The path to ubiquitous AI (17k tokens/sec)

#201
Holy Hallucinations, Batman...

User You: What is the airspeed of an unladen swallow/

Jimmy:

A reference to the classic line from Monty Python and the Holy Grail!

Thus far it is correct but here it comes:

In the movie, a group of medieval scholars, the Knights who say "Ni!" (also known as the Bridgekeepers), ask a king's servant, a blacksmith, if an unladen swallow makes good time. The blacksmith, played by John Cleese, responds with the now-famous line:

"And what's the airspeed velocity of an unladen swallow?"

To which the blacksmith responds:

"What do you mean? An unladen swallow?"

The blacksmith then explains that an unladen swallow is when a swallow (a type of bird) is flying in the air without carrying any burden, implying that it is moving at a speed that is not affected by carrying any weight or load.

Nope, that is not at all how it went. No blacksmiths, not related to the Knights who say 'Ni!' - who have nothing scholarly about them, they wanted a shrubbery and demanded another one when their wish was fulfilled. It was the Bridgekeeper - not related to the aforementioned knights - who asked 3 questions and would throw anyone who could not answer all three into the chasm. When he asked Arthur - played by Cleese - the question about the airspeed of the swallow he asked the Bridgekeeper whether he meant an African or European swallow. The Bridgekeeper did not have an answer and was thrown into the chasm, problem solved.

However, in reality, swallows are not typically "unladen" in flight. They are small birds that fly at relatively fast speeds, usually around 10-15 km/h (6-9 mph), but they are not usually carrying any burdens!

Needless LLM-blabber.

The "airspeed velocity of an unladen swallow" has become a meme and a cultural reference point, often used humorously or ironically to refer to situations where someone is trying to make an absurd or non-sensical argument or ask an absurd question.

Somewhat correct but not necessary in this context.

The correct answer to the question would have been Do you mean an African or European swallow? followed by a short reference to the movie.

Of course this demo is not about the accuracy of the model - 'an old Llama' as mentioned elsewhere in this thread - but it does show that speed isn't everything. For generating LLM-slop this hardware implementation probably offers an unbeatable price/performance ratio but it remains to be seen if it can be combined with larger and less hallucination-prone models.

Re: The path to ubiquitous AI (17k tokens/sec)

#202

This is really cool! I am trying to find a way to accelerate LLM inference for PII detection purposes, where speed is really necessary as we want to process millions of log lines per minute, I am wondering how fast we could get e.g. llama 3.1 to run on a conventional NVIDIA card? 10k tokens per second would be fantastic but even at 1k this would be very useful.

PII redaction is a really good use-case.

Also, "10k tokens per second would be fantastic" might not be sufficient (even remotely) if you want to "process millions of log lines per minute".

Assuming a single log line at just 100 tokens, you need (100 * 2 million / 60) ~ 3.3 million tokens per second processing speed :)

Re: The path to ubiquitous AI (17k tokens/sec)

#203
post #126

Earlier quoted context omitted.

This math is useful. Lots of folks scoffing in the comments below. I have a couple reactions, after chatting with it: 1) 16k tokens / second is really stunningly fast. There’s an old saying about any factor of 10 being a new science / new product category, etc. This is a new product category in my mind, or it could be. It would be incredibly useful for voice agent applications, realtime loops, realtime video generati…

For speculative decoding, wouldn’t this be of limited use for frontier models that don’t have the same tokenizer as Llama 3.1? Or would it be so good that retokenization/bridging would be worth it?

I think they’d commission a quant directly. Benefits go down a lot when you leave model families.

Re: The path to ubiquitous AI (17k tokens/sec)

#206
post #151

Earlier quoted context omitted.

I'm wondering how much the output quality of a small model could be boosted by taking multiple goes at it. Generate 20 answers and feed them back through with a "rank these responses" prompt. Or doing something like MCTS.

Isn't this what thinking models do internally? Chain of thoughts?

No. Chain of thought it just the model generating a single answer for longer inside tags which are not shown in the final response. The strategy of generating different answers in parallel is something different (which can be used in conjunction with chain of thought) and is the thing used by models like Gemini 3 Deep Think and GPT-5.2 Pro.

Re: The path to ubiquitous AI (17k tokens/sec)

#207

Holy cow their chatapp demo!!! I for first time thought i mistakenly pasted the answer. It was literally in a blink of an eye.!! https://chatjimmy.ai/

I asked, “What are the newest restaurants in New York City?” Jimmy replied with, “2022 and 2023 openings:” 0_0

Well, technically it's answer is correct when you consider it's knowledge cutoff date... it just gave you a generic always right answer :)

Re: The path to ubiquitous AI (17k tokens/sec)

#208

Jesus, it just generated a story in 0.039s. Whoever doesn’t buy/replicate this in the next year is dead. Imagine OpenAI trying to sell you a platform that takes 15 minutes, when someone else can do it in 0.001s.

> generated a story Was the story any good? The model used isn't comparable to something OpenAI would offer you today. What good is a fast gibberish generator?

Question:

> Generate a fiction book about an old sorcerer who, like all elderly magicians, has to go back to studying magic for two years at a magic school in order to be granted the right to keep his wand and to keep doing magic at his home.

Answer, generated in 0,063s @ 15 742 tok/s:

> The Relearning of Eryndor Thorne

> In the quaint village of Brindlemark, where ancient traditions and mystical energies coexisted, Eryndor Thorne had lived a life of quiet reverence. For nigh on five decades, he had wielded the arcane arts with precision and mastery, earning him the respect of his peers and the awe of the villagers. His home, a humble abode atop a hill overlooking the rolling hills and verdant valleys, was a sanctuary of sorts, a place where the very fabric of magic was woven into the air. His library was a treasure trove of ancient tomes, texts, and scrolls, filled with the wisdom of the ages. Yet, despite his mastery over the arcane, Eryndor felt the stirrings of a nagging concern. The once-renewable license to practice magic, a privilege granted by the Council of Mages, was on the cusp of expiration.

> As all elderly mages knew, every two decades, the Council would demand that those who had spent a lifetime mastering the arcane arts return to the prestigious Wysteria Academy, the hallowed halls of magic learning, to relearn the foundational principles of magic. This was a ritual as old as the craft itself, a test of the elderly mage's understanding, to prove they still possessed the fundamental knowledge and skills required to practice magic with precision and safety.

> Eryndor, like many of his fellow mages, had grown complacent in his mastery. He had mastered the intricacies of elemental magic, bending the winds, waters, and flames to his will. His wand, a family heirloom passed down through generations, had become an extension of his arm, channeling the power of magic with ease. Yet, despite his proficiency, the Council's decree was clear: Eryndor must return to Wysteria Academy for two years of reeducation.

> The journey was not an easy one. Packing his worn leather satchel with the few books and scrolls he deemed essential, Eryndor bid farewell to his loved ones in Brindlemark, leaving behind the familiarity of his home to embark on a journey that would take him to a place both wondrous and treacherous. At Wysteria Academy, he joined a cohort of students from all corners of the realm, each as skilled and eager as he once was. The relearning process was a grueling one, as if facing a mountain he had once conquered. New techniques were presented, new theories to unravel, and new expectations to meet. Eryndor, once a master, now sat at the feet of his teachers, absorbing the new knowledge like a parched traveler at an oasis.

> Among his fellow students, there was the young, fire-kissed mage, Elara, who wielded magic with an intensity that bordered on reckless abandon. Her fiery nature and quick wit often put her at odds with the strict, ancient traditions, earning her a certain notoriety among the academy's elder mages. Then there was the enigmatic, shadow-drawn Kael, whose mastery of the arcane was matched only by his mystery. Kael's affinity for the dark arts raised more than a few eyebrows among the faculty, but Eryndor, having once walked the fine line between light and shadow, saw something of himself in the young mage.

> As the years passed, Eryndor grew to appreciate the challenges and opportunities that came with his return to the academy. He found himself grappling with the nuances of magic anew, rekindling memories of his early days as a novice. The relearning process was as much about rediscovering himself as it was about mastering the arcane. His studies were a journey of self-discovery, one that tested the mettle of his will and the depths of his understanding.

> Upon completion of his studies, Eryndor stood before the Council once more, his wand in hand, his heart afire with the thrill of rediscovery. The Council's examination was not merely a test of his knowledge but a test of his character. Eryndor, like many of his peers, had grown complacent, but the rigors of relearning had rekindled a spark within him, a flame that would guide him through the trials ahead.

> With his renewed license granted, Eryndor returned to Brindlemark, his home and his heart rejuvenated. His library, once a testament to his mastery, was now a canvas for the new knowledge he had acquired. His wand, now an extension of his rekindled passion for the arcane, channelled magic with a precision and flair that was unmatched. Eryndor Thorne, the elderly mage, had not only relearned magic but had rediscovered himself, a testament to the transformative power of education and the enduring spirit of the arcane.

Re: The path to ubiquitous AI (17k tokens/sec)

#209

Earlier quoted context omitted.

Isn't this what thinking models do internally? Chain of thoughts?

No. Chain of thought it just the model generating a single answer for longer inside tags which are not shown in the final response. The strategy of generating different answers in parallel is something different (which can be used in conjunction with chain of thought) and is the thing used by models like Gemini 3 Deep Think and GPT-5.2 Pro.

Hmm.. got it. Thanks..

Re: The path to ubiquitous AI (17k tokens/sec)

#210
post #21

This is not a general purpose chip but specialized for high speed, low latency inference with small context. But it is potentially a lot cheaper than Nvidia for those purposes. Tech summary: - 15k tok/sec on 8B dense 3bit quant (llama 3.1) - limited KV cache - 880mm^2 die, TSMC 6nm, 53B transistors - presumably 200W per chip - 20x cheaper to produce - 10x less energy per token for inference - max context size: flexib…

There’s a bit of a hidden cost here… the longevity of GPU hardware is going to be longer, it’s extended every time there’s an algorithmic improvement. Whereas any efficiency gains in software that are not compatible with this hardware will tend to accelerate their depreciation.
Post reply on HN