Tech | Spirituality | Fitness | Humor | Current Trends | 🇮🇳 First
Joined May 2009
- Tweets17.6K
- Following348
- Followers653
- Likes19.9K
Harness design
Agent = Model + Harness.
The model is the brain. The harness is everything you build around it: the tools, the memory, the guardrails and the checks that run before it says "done".
The same model in two different harnesses gives very different results. That's a big part of why coding tools built on similar models can feel so different to use.
The habit that matters most: when the agent makes a mistake, don't just rewrite the prompt and hope. Change the system.
If it keeps pushing code that fails tests, don't write "please be careful". Add a rule that blocks the commit until the tests pass.
Every failure becomes a permanent fix, and reliability compounds.
1
6
The first list changes every few months. New versions, new names, new leaderboards. Half of it will look dated by next year.
The second list hasn't changed since agents began, and it applies to every model on the first one.
Learn tokens, context, tools, evals and harness design, and you can pick up any new model in a weekend.
Learn only the model names, and you start over with every release.
1
4
Attending a day-long training on Microsoft Foundry.
Have been working on open source tools for some time now.
15
Anthropic just released two new AI models, Claude Fable 5.2 and Claude Mythos 5.2, and the details say a lot about where AI is heading.
Claude Fable 5.2 is built for coding, reasoning and AI agent tasks. Anthropic says it's the one to reach for when the problem is hard, like engineering, mathematics and robotics. For everyday work, the company points to Claude Opus 5.5, which it released in September. So the lineup is getting clearer: Opus for daily tasks, Fable for the toughest problems.
Claude Mythos 5.2 is a different kind of model. It's aimed at approved cybersecurity and life sciences research, which means it isn't something most people will use directly. Fable is the publicly available version of Mythos, with additional safeguards added on top.
Both new models generate output up to 30% faster than their 5.1 predecessors, and Anthropic says they communicate more clearly with users. They're priced at $8 per million input tokens and $40 per million output tokens, which is lower than the pricing of earlier Fable and Mythos releases. So you get a faster model that costs less, which is exactly what developers building agents have been asking for, since agents burn through a lot of tokens.
The story behind the launch matters too. These are Anthropic's first frontier models since CEO Dario Amodei called for slowing down the pace of AI development back in September. The company says both models went through rigorous safety and alignment testing, and that it built new evaluations and post-launch monitoring tools specifically for this release. In other words, the message is that speed and caution can go together.
The timing is hard to miss as well. Anthropic is reportedly preparing for an IPO that could come as soon as next week, although that hasn't been confirmed by the company. It also rolled out Claude Haiku 5.5 in October, so it has been shipping steadily while competition heats up. OpenAI, Google and Meta are all pushing out new models focused on coding, reasoning and agents, because those are the areas enterprises are willing to pay for.
Models are getting faster and cheaper at the same time. The market is splitting into everyday models and frontier models for hard problems. And safety work is becoming part of the launch itself, not an afterthought.
Benchmarks and press releases only tell part of the story. The real test is what developers build with these models in the next few weeks.
1
2
148
When Lord Rama’s patience turned into apocalyptic fury. 🔥🏹
In the Valmiki Ramayana, Sri Rama is revered as "Jitakrodhah" (the master of his anger). But he wasn’t an emotionless statue; he was a human embodiment of Dharma who experienced raw, intense, and righteous fury when pushed to the absolute brink.
The peak of this wrath happens in the Aranya Kanda (Sarga 64), right after he discovers clues of Sita’s violent abduction. Believing the entire cosmos stood by silently and allowed a sin to occur, his grief instantly transforms into catastrophic rage.
Valmiki describes Rama’s eyes flushing copper-red as he grips his bow tightly and draws an arrow blazing like a venomous serpent.
Here is the complete sequence of verses (72 to 76) capturing this terrifying moment:
नाशयामि जगत्सर्वं त्रैलोक्यं सचराचरम् ।
इत्युक्त्वा रोषताम्राक्षो रामो निष्पीड्य कार्मुकम् ॥ ७२ ॥
(Saying, "I shall destroy the entire universe, including all three worlds with all their mobile and immobile beings," Rama, his eyes turning copper-red with intense fury, firmly gripped his bow.)
शरमादाय सन्दीप्तं घोरमाशीविषोपमम् ।
सन्धाय धनुषि श्रीमान् रामः परपुरञ्जयः ॥ ७३ ॥
(Taking a blazing, terrible arrow that resembled a deadly, venomous serpent, the glorious Rama, conqueror of enemy cities, fixed it firmly to his bowstring.)
युगान्ताग्निरिव क्रुद्ध इदं वचनमब्रवीत् ।
यथा जरा यथा मृत्युर्यथा कालो यथा विधिः ॥ ७४ ॥
(Blazing with wrath like the all-consuming fire at the end of an epoch, he spoke these words to Lakshmana: "Just as old age, death, time, and destiny can never be counteracted by any living creature at any time...")
नित्यं न प्रतिहन्यन्ते सर्वभूतेषु लक्ष्मण ।
तथाऽहं क्रोधसंयुक्तो न निवार्योऽस्मि सर्वथा ॥ ७५ ॥
("...likewise, O Lakshmana, when I am unified with this fury, I am absolutely unstoppable by anyone or anything in any way whatsoever.")
पुरेव मे चारुदतीमनिन्दितां दिशन्ति सीतां यदि नाद्य मैथिलीम् ।
सदेवगन्धर्वमनुष्यपन्नगं जगत्सशैलं परिवर्तयाम्यहम् ॥ ७६ ॥
("If that flawless princess of Mithila, who has beautiful teeth and a charming smile, is not returned to me today in the exact fine condition she was in before, I will completely upturn this entire universe, along with its gods, celestial musicians, humans, serpents, and mountains.")
What makes Valmiki’s portrayal so profound is this psychological realism. Rama’s anger wasn't born out of toxic malice, but out of a deep sense of betrayal that his standard gentleness (Kshama) was being misconstrued as weakness.
2
69
You're charged for AI by a unit most people have never looked at.
It isn't words. It isn't characters. It's tokens, and understanding them explains why some prompts are slow, some answers are expensive, and some models feel snappier than others.
Here’s why 👇🏼
🤖 Made with AI
1
6
85
So what does "token efficiency" really mean?
Saying the same thing in fewer tokens. Not resending what the model already has. Keeping the context tight so attention and the cache stay small. Reusing cached prefixes. Using a smaller model for simple tasks.
Done well, you get lower cost, lower latency and often better answers, because the model isn't wading through noise.
1
1
13
A worked example. Same request, two ways.
Version A (142 characters, 27 words):
"Could you please be so kind as to provide me with a summary of the following article, if at all possible, in around three short bullet points?"
Version B (61 characters, 9 words):
"Summarise the following article in three short bullet points."
Using the rough rule of four characters per token, A is about 35 tokens and B is about 15. Exact counts depend on the model's tokenizer.
Same task, same result, and well over half the tokens gone. Now imagine that across thousands of requests a day.
2
15