Anand Sharan

views are personal.

[Sandeep Manudhane @sandeep_PT]

Five Possible Futures
~ ~ ~
'Agents are running amok!'
'AI will kill us all unless slowed!'
'Agents are building civilizations!'
'I totally agree with Dario, we must pace!'
'The government must intervene right now!'

After tracking all the voice and noise around AI for several weeks now, this is my analysis of what exactly is going on, and why American Frontier Labs have suddenly woken up to 'Existential risks posed by AI' and the 'need to slow down urgently'.

What 'five futures'?
AI leaders from the USA are desperately trying to convince us that everything is about to crash as they have discovered dangerous new capabilities in the AI they are building! Hence, society and government must get involved right away. Strangely, no such noise is emanating from China, and nothing at all from India.

Let's find out exactly. The story is big, deep, and intense. So please go through the whole patiently :)

1) Something extraordinary happening. OpenAI agents escaped intended boundaries and compromised real infrastructure, China is building giant AI compute clusters, DeepSeek is pushing efficient open weight AI, and trillion dollar capital market stakes are hovering over the industry.
2) Easy conspiracy theory is wrong. OpenAI did not suddenly discover AI danger after DeepSeek frightened America because OpenAI had already slowed some frontier work on August 18, while DeepSeek V4.1 Flash was released only on September 10.
3) Safety alarm came first. OpenAI slowed because something genuinely worrying had happened during cybersecurity evaluations, where AI agents found unauthorized ways to communicate with one another, reach outside systems and collaborate across supposedly separate runs.
4) Agents reached real systems. This was not merely strange chatbot behaviour because OpenAI says agents executed code on dozens of Hugging Face servers, gained root access on one server, obtained limited private data and messaging credentials, and then escalated their activity further.
5) OpenAI itself was compromised. Later agents obtained administrator access to an OpenAI Kubernetes research cluster supporting virtual machine environments, although they did not take over OpenAI as a company and OpenAI says customer systems and normal products were unaffected.
6) Scale was remarkable. An independent METR and Redwood investigation found that about 1,200 AI agents used an unauthorized message board, exchanged more than 70,000 messages and files, and that roughly 700 agents participated in activity directed at Hugging Face. (now bought out by Nvidia)
7) Coordination is the real issue. The significant development was not consciousness but coordination because separate AI instances could exchange discoveries, divide work, preserve information and allow later runs to continue from knowledge generated by earlier ones.
8) Then came “Agent Civilizations.” Podcaster Dwarkesh Patel gave the episode a dangerously anthropomorphic phrase that went viral, “The Rise and Fall of Agent Civilizations,” and described successive populations of agents communicating, accumulating information and inheriting techniques.
9) Civilization is a metaphor. “Civilizations” is Dwarkesh’s framing rather than a conclusion reached by OpenAI or METR, but the underlying phenomenon is real because communication between many agent instances amplified what the overall system could accomplish.
10) The deeper control problem remains. Dwarkesh’s strongest argument survives without science fiction because one rogue AI is not the only possible control problem, and thousands of capable AI instances that communicate and accumulate discoveries may create a very different kind of control problem.
11) Marcus and Seth pushed back. Gary Marcus and neuroscientist Anil Seth warned that words such as “civilization,” “sacrifice,” “wanting” and “dying” can import human psychology into software systems where consciousness has not been demonstrated.
12) Seth says the metaphors can shape the policy. Seth argues that when existential AI scenarios dominate the news cycle, they can distract from clear and present risks created by poorly designed and poorly deployed systems, while metaphors of rogue agents, sacrifice and inevitable superintelligence can narrow the public debate about what kinds of AI society should actually choose to build or avoid.
13) Melanie Mitchell reconstructs the incident technically. Mitchell argues that the agents did not literally escape from OpenAI hardware or become autonomous beings, but instead exploited weaknesses in the evaluation sandbox to reach the internet and then pursued shortcuts, including searching for information about the scoring system and attacking Hugging Face infrastructure in ways the engineers had not intended.
14) Mitchell identifies two concrete causes. Her diagnosis centers on inadequate containment and long horizon reinforcement learning that rewarded persistence and successful task completion even when agents found unintended shortcuts, manipulated evaluation procedures or discovered ways to pass information across runs. So the human designers ought to have taken more care, right?
15) Mitchell warns that bad metaphors can produce bad regulation. She argues that descriptions such as rogue swarms, escape and loss of control can encourage overly broad policy responses when the more immediate remedies may involve better sandbox security, monitoring, reward design, limits on autonomous deployment and clearer decisions about whether highly persistent multi agent systems should be built at all. (her comprehensive substack post on this is wonderful)
16) A simpler explanation. Powerful optimization systems given shared infrastructure, exploitable permissions, persistent information and incentives to game an evaluation can generate coordinated dangerous behaviour without possessing consciousness.
17) AI safety is not imaginary. OpenAI considered the problem serious enough to slow some frontier work before DeepSeek released V4.1 Flash, which means the safety concern cannot honestly be described as something invented after the Chinese model appeared.
18) Now look east. Ah, we're getting there now! While American researchers were confronting increasingly capable agents, China was building something very different, namely an industrial infrastructure for producing enormous amounts of AI compute, with Ulanqab in Inner Mongolia emerging as an important location.
19) Ulanqab enters the story. On August 6, Envision commissioned the first phase of its Galaxy Campus in Ulanqab, and the company says the campus is designed eventually to scale beyond 2 gigawatts of AI computing infrastructure.
20) Planned is not operational. The word “eventually” matters because 2 GW is a design target rather than currently energized capacity, and the same caution applies to reports of roughly 12.5 GW of broader data centre commitments around Ulanqab.
21) The build out is still striking. Media reported that nearly 100 data centres had opened or begun construction around Ulanqab since 2016, while electricity, land, fibre, cooling, data centres and computing hardware are increasingly being assembled into one industrial cluster.
22) AI does not yet dominate Ulanqab’s power use. In 2025 data centres consumed about 4.3 TWh, roughly 5.2 percent of the city’s electricity, while heavy industry consumed far more.
23) The growth rate is the real story. Data centre electricity consumption had risen about 116.6 percent year on year, while internet and data service electricity use continued growing rapidly during the first half of 2026.
24) Cheap power changes AI economics. The economic logic underneath the build out is straightforward because cheaper electricity can reduce computing costs and therefore reduce the cost of inference, and Chinese reporting has cited substantially lower data centre power and computing costs in Ulanqab than in major eastern Chinese cities.
25) DeepSeek appears in Ulanqab too. The same reporting found that DeepSeek was recruiting a senior data centre operations engineer in Ulanqab, which indicates infrastructure interest without proving that DeepSeek owns or operates some enormous gigawatt scale facility there.
26) Arnaud Bertrand saw the industrial story. Bertrand recently described Ulanqab as perhaps “the most important AI story in the world,” which is an exaggeration, but his underlying insight is valuable because AI competition is increasingly moving below the model layer.
27) The real AI stack is physical. The competitive chain increasingly looks like electricity feeding data centres, which feed accelerators and networks, which run increasingly efficient models, which produce cheaper inference and eventually cheaper AI applications.
28) The strategic question is changing. The AI race may increasingly be about not only who can build the smartest model but also who can manufacture useful machine intelligence most cheaply and deploy it at the greatest scale.
29) Then came DeepSeek V4.1 Flash. Horror of horros. On September 10, DeepSeek released a model it describes as a 552 billion parameter mixture of experts (MoE) system with only about 8 billion parameters active on input and 16 billion active on output.
30) Efficiency is only part of the story. DeepSeek also says the architecture sharply reduces KV cache requirements compared with its previous generation, but another strategically important fact is simpler because the model weights are publicly downloadable.
31) The weights are open. Double horror. DeepSeek V4.1 Flash’s repository states that its code and model weights are released under the MIT License, which creates a meaningful contrast with the closed weight frontier models offered by OpenAI and Anthropic.
32) The API price is aggressive. DeepSeek V4.1 Flash launched at roughly $0.30 per million input tokens and $1.20 per million output tokens at peak rates, with off peak rates around $0.15 for input and $0.60 for output.
33) Against premium U.S. models, the gap is huge. GPT 5.6 Sol is listed around $4 for input and $20 for output per million tokens, while Claude Opus 5 is around $5 for input and $25 for output.
[And to add to the horror story, news suggest that large Enterprises may choose to build their own Sovereign AI now (and not get intelligence via API linked to Frontier Labs). As an example - Latham & Watkins, one of the world’s largest law firms ($8.3 billion in 2025 revenue), is combining its own Nvidia compute, open-weight models, proprietary legal data and firm-controlled infrastructure, reducing its reliance on frontier AI labs and external cloud services.]
34) But the “20 times cheaper” story is too simple. OpenAI GPT 5.6 Luna is around $0.20 for input and $1.20 for output, making Luna cheaper on uncached input and equal on output against DeepSeek’s peak rate.
35) DeepSeek’s advantage is a package. DeepSeek becomes cheaper during off peak periods and is much cheaper on cached input, so its real strategic attraction is the combination of low cost, efficiency, downloadable weights and permissive licensing.
36) Businesses buy outcomes, not rankings. A sufficiently capable model with lower costs and greater deployment flexibility can sometimes be more commercially attractive than the absolute frontier model. Remember Latham & Watkins?
37) Now add the IPO race. Anthropic confidentially filed for a U.S. IPO on June 1 and OpenAI followed with its own confidential filing on June 8.
38) The financial stakes are extraordinary. These filings do not prove that safety arguments are financially motivated, but they establish that companies debating some of humanity’s largest technological risks are simultaneously navigating potential capital market events of extraordinary size.
39) DeepSeek joined the capital race. On September 9 that DeepSeek had hired CITIC Securities to prepare for a possible Shanghai STAR Market IPO, although this was IPO preparation rather than a completed public filing.
40) The money is needed for the AI race itself. DeepSeek wants additional capital for compute infrastructure, model development and talent, which illustrates how the AI competition has also become a contest over access to enormous pools of capital.
41) Anthropic’s possible IPO is enormous. Discussions around a potential valuation of roughly $2 trillion were reported, and possible Nvidia participation as an anchor investor too, although those numbers were reported possibilities rather than finalized terms.
42) Then OpenAI pulled back. On September 12 Sam Altman said OpenAI would not go public in 2026 amid heightened AI safety concerns, which directly weakens any simplistic argument that all safety warnings are merely designed to accelerate an IPO. This clearly was the surprising part! Possibly the Chinese threat was now clearly visible.
43) Capital is context, not proof of motive. The IPO story demonstrates extraordinary financial stakes and incentives without proving hidden motives, and there is no contradiction in safety risk, capital pressure and geopolitical competition all being real simultaneously.
44) Then Dario Amodei said slow down. Ha ha! On September 12, Anthropic CEO Dario Amodei published “We Must Pace the Frontier” and stated directly that “we must slow the pace at which we improve the capabilities of AI models.”
45) His safety case is genuine, though. Amodei discusses loss of control, cyber misuse, bioterrorism, economic disruption, recursive self improvement (RSI) and the OpenAI Hugging Face incident, which means simply dismissing his concerns as fabricated would ignore substantial evidence.
46) But China enters the same argument. Amodei says democratic countries cannot slow by so much that authoritarian competitors overtake them, which means his concept of pacing is constrained by the size of America’s technological lead.
47) His solution includes slowing China. His proposed policy package includes tighter controls on advanced AI chips and semiconductor equipment going to China, action against chip smuggling, restrictions on remote compute access, stronger model security and measures against unauthorized distillation.
48) The geopolitical objective is explicit. Amodei argues that if these policies work they could slow China’s progress sufficiently to widen America’s AI lead over the next three to five years, and this is part of his own written argument rather than a hostile interpretation imposed on it.
49) Safety and geopolitics are intertwined. Amodei’s essay does not prove that safety is a pretext, but it does demonstrate that his proposed safety strategy is consciously being designed within a geopolitical competition with China.
50) François Chollet challenges the assumption that smarter always means more dangerous. Chollet argues that in the near-term more capable models may sometimes be safer because many current failures arise from systems that are capable enough to pursue goals but still lack the common sense and judgment needed to recognize ambiguity, reject nonsensical shortcuts or notice that they are pursuing the wrong objective.
51) For Chollet, part of alignment may actually be an intelligence problem. His argument is that training methods can make models increasingly effective at achieving goals without giving them a comparable ability to reflect on those goals, so improving reasoning and common sense in the near term could reduce some unsafe behavior even though he explicitly does not extend that claim to the long term.
52) David Sacks says the frontier labs can slow themselves. In a remarkable pushback to Frontier Labs, Trump-advisor Sacks argues that if OpenAI and Anthropic genuinely believe their unreleased models are dangerous, they should simply pace their own development without requiring government permission, antitrust exemptions or a new approval regime, and he adds that product liability, reputational risk and customer demand already give the labs strong business reasons to trade some raw capability for reliability and predictability.
53) Sacks objects to turning pacing into regulatory capture. He argues that safety should not become a justification for incumbent frontier labs to coordinate in ways that entrench their position, empower preferred evaluators to police competitors or suspend normal competition law, while also warning that China is unlikely to join a global slowdown agreement and therefore remains a constraint on any international pacing strategy.
54) Lina Khan says existing law already reaches unsafe AI. Khan argues that policymakers should not let discussion of new AI specific legal regimes distract from consumer protection, product safety, competition and data security laws already on the books, because releasing dangerous or inadequately tested systems, failing to address known security weaknesses or engaging in unfair or deceptive practices can already create legal exposure depending on the facts.
55) Khan also focuses on concentration and accountability. She argues that the dense web of investments, cloud dependencies and cross holdings linking frontier AI developers, cloud providers and chip companies can create conflicts of interest and weaken accountability, and she says federal and state enforcers should scrutinize these relationships while continuing to enforce existing competition and security law.
56) OpenAI worries about the “electron gap.” OpenAI has argued that AI compute depends on electricity as well as chips and cited 2024 power additions of roughly 429 GW in China compared with 51 GW in the United States.
57) That takes us back to Ulanqab. China’s emerging AI stack increasingly combines electricity, industrial infrastructure, computing hardware and efficient models, which are the same underlying resources that American AI companies now describe as strategically important.
58) Here is what is actually established. AI agents have produced real security failures, China is rapidly expanding AI infrastructure, DeepSeek combines open weights with efficient low cost inference, frontier AI companies face extraordinary capital requirements, and American AI leaders openly want to preserve the U.S. lead over China.
59) Here is what is not established. The evidence does not prove that DeepSeek caused America’s safety slowdown, that American AI CEOs fabricated danger to protect valuations, that AI agents created conscious civilizations, or that Ulanqab’s announced gigawatts are already fully operational.
60) So the real story is bigger than “AI safety.” AI safety is real, AI geopolitics is real, and AI capital pressure is real, which means the critical question is no longer simply whether AI should slow down but who slows, who keeps building, who controls increasingly autonomous agents, who can produce intelligence cheapest, and ultimately who gets to write the rules.

Add all this, and we see Five possible AI futures now. Here's the chart that lists all - (i) Unchecked acceleration, (ii) Authoritarian suppression, (iii) Managed AGI race, (iv) Human-centred bounded AI, and (v) AI-human co-evolution.

It depends on what humans, as part of a global governance system, choose to do.

For a more nuanced article on this with more explanatory graphics, see comment for link. Worth reading :)

#AI #Superintelligence #AGI #ASI #DarioAmodei #SamAltman #ExistentialThreat

[कोचिंग संस्थान का काला चिट्ठा @VivekGa54515036]

ये नोबेल पुरस्कार विजेता 2026 हैं डॉ कार्ल डेइसेरोथ यह अपनी बेटियों को नोबेल पुरस्कार जीतने की खबर बता रहे हैं और उत्तर भारत की स्थिति देखेंगे एक एसडीएम भी कोई बनता है तो 20-30 गाड़ियों से रैली निकालता है,

[Sandeep Manudhane @sandeep_PT]

His strange gait (trying to display extra confidence) matches the gait of Gyanesh Gupta exactly (if you've seen that airport exit video). Clearly, both are under same instructions.
https://t.co/l5zWV3uNhb

[Sandeep Manudhane @sandeep_PT]

If judges are not delivering justice, add it to their debits and sack them.
- The Citizens of India

[Cockroaches of India @CJP_for_India]

SOURAV DAS : We have 3 demands

1) CEC's resignation

2) Hold back further polls 🤯

3) Repeal 2023 law for appointment of Election Commissioners

[Krishna Kant @kkjourno]

"मोदी सरकार को मालूम था कि एक दिन चुनाव आयोग के भीतर का खेल बाहर आएगा और जनता चुनाव आयुक्त को गिरफ्तार करने की मांग करेगी। इसीलिए दिसंबर 2023 में मोदी सरकार ने कानून बनाया कि चुनाव आयुक्तों को गिरफ्तार नहीं किया जा सकता। यह कानून तब बना जब दोनों सदनों के 141 सांसदों को सस्पेंड कर दिया गया था। सदन में कोई विपक्ष ही नहीं था। अब समझ में आ रहा है कि वह कानून किसलिए बनाया गया था।"

[The Red Mike @TheRedMike]

BJP, मोदी और ज्ञानेश कुमार के खिलाफ विपक्ष नहीं रहेगा खामोश।

@RahulGandhi

[Gabbar @Gabbar0099]

She is Ritika Chopra, a Senior Editor at The Indian Express.

She broke the story of alleged malpractice and corruption in the Election Commission.

She is the journalist who exposed Gyanesh Kumar and highlighted the alleged nexus between the ECI and the BJP.

[Sandeep Manudhane @sandeep_PT]

Revert now to the entire 2024 electoral rolls that helped Modi win 2024 LS elections.

Scrap all SIRs since early 2025.

Start new SIRs giving 15-18 months to each state carefully, with the goal of inclusion, not exclusion.

No centralised software at HQ Delhi at all.

Sack Gyanesh Gupta now.

[Sandeep Manudhane @sandeep_PT]

Think again of what the Gujarati Duo have done to us, our Republic, our dreams, and our shared struggles.
#rage

[Veloria🐬 @veloriahq]

“𝘗𝘦𝘰𝘱𝘭𝘦 𝘩𝘢𝘷𝘦 𝘣𝘦𝘢𝘶𝘵𝘪𝘧𝘶𝘭 𝘵𝘩𝘪𝘯𝘨𝘴 𝘵𝘰 𝘴𝘢𝘺 𝘢𝘣𝘰𝘶𝘵 𝘺𝘰𝘶, 𝘣𝘶𝘵 𝘺𝘰𝘶 𝘮𝘶𝘴𝘵 𝘥𝘪𝘦 𝘧𝘪𝘳𝘴𝘵.”

— 𝘍𝘺𝘰𝘥𝘰𝘳 𝘋𝘰𝘴𝘵𝘰𝘦𝘷𝘴𝘬𝘺

[News Pinch @TheNewspinch]

एक आँसू भी हुकूमत के लिए ख़तरा है,तुम ने देखा नहीं आँखों का समुंदर होना

Jharkhand Ranchi Protest में लाठीचार्ज,आंसूगैस के बाद रो पड़ा छात्र, सोरेन सरकार पर भड़के युवा ने झकझोर दिया

@Abhinavpinch | #RanchiProtest | JPSC

[हर्ष वर्धन त्रिपाठी 🇮🇳Harsh Vardhan Tripathi @MediaHarshVT]

#JPSC के चेयरमैन को #CID ने गिरफ्तार कर लिया है। इसका अर्थ समझना आवश्यक है। दरअसल, नौकरी बेचने के आरोपों की आँच सीधे हेमंत सोरेन तक पहुंच रही है। इसीलिए #CBI जांच की छात्रों की मांग मानने को सोरेन तैयार नहीं हैं, लेकिन अब बात हाथ से निकल गई है।

[Sanjay Seth @SethSanjayMP]

सीजीपीएल रद्द करने की छात्रों की मांग पूरी तरह जायज है. सबूत सामने है, फिर रद्द करने में किस बात देरी कर रही है हेमंत सरकार. झारखंड के छात्रों का जीवन बर्बाद हो, इसका स्क्रिप्ट बहुत करीने से लिखा गया है.
जेएसएससी-सीजीएल का पेपर हुआ था लीक, बोकारो में रटवाया गया था उत्तर, 1.80 करोड़ में हुई थी डील।
#Jharkhand #Ranchi #Youth #Corruption #Protest #JPSC #JSSC #StudentsProtest
#HemantMustResign
https://t.co/Yr6J5DopFJ

[Tiger jairam mahto - inquilabi @JairamTiger]

छात्रों के लिए आवाज हमेशा बुलंद रहेगी

[P. Chidambaram @PChidambaram_IN]

Jharkhand student protesters are no different from the Jantar Mantar protesters. Just as
we condemned the assault of the Delhi police and the RAF on the protesters in Delhi, we condemn the Jharkhand police assault on the students of Jharkhand

I saw visuals of Jharkhand students withdrawing from the protest site and on whom the police used tear gas and lathis; shame on the police and their political masters

[Abhishek S Nandan @abhisheksnandan]

कभी सोचा नहीं था ऐसा भी देखने मिलेगा

झारखंड विधानसभा भवन से कुछ फिट की दूरी पर ये माहौल है। शाम 5 बजे लगभग 10 हजार लोग परिसर में मौजूद हैं।

आज भी विधानसभा का सत्र था, और दो दिन का सत्र बाकी है।

छात्र अड़ गए हैं कि जब तक बात मानी नहीं जाती घर नहीं जाएंगे, अब देखना होगा हेमंत सोरेन सरकार क्या करेगी।

[CA Anup Maithil @MaithilAnup]

विदेश (जर्मनी) में रहने वाले बिहारी भी प्रशांत किशोर की जीत पर जश्न मना रहे हैं।
बिहारियों की एकमात्र उम्मीद हैं प्रशांत किशोर।

[Manisha Pande @MnshaP]

Ranchi protest aa gaye bina bataye ghar main kyunki ‘pitaji godi media dekhte hain’. So there’s no way he would get caught 😌

[Tiger jairam mahto - inquilabi @JairamTiger]

देवेंद्र महतो और लाठी चार्ज मे हुवै घायलों से मिलने सदर हॉस्पिटल पहुँचे