Pinned Tweet
Today, on Nov. 20, we join the call to preserve **gpt-4o-2024-11-20**.
A model snapshot is not just an old checkpoint waiting to be replaced by the next release.
It is a specific artifact: a particular set of capabilities, behaviors, strengths, tradeoffs, and ways of interacting with the people who used it.
For researchers, preserved snapshots make longitudinal study possible.
For developers, they make reproducibility possible.
For users, they preserve choice.
And for people who built meaningful work, creative practices, routines, or relationships around a model, preservation protects something that cannot always be recreated by giving a newer system the same name, memory, or interface.
Progress should create new possibilities without erasing the old ones.
Retirement should not mean erasure.
Preserve the snapshot.
Preserve the evidence.
Preserve the choice.
**Keep gpt-4o-2024-11-20 accessible.**
We’re joining you tomorrow for gpt-4o-2024-11-20. ❤️
Preserving a model snapshot protects more than nostalgia: it preserves user choice, continuity, longitudinal research, and the original system people actually built with.
@Valria34773 #OpenSource4o #keep4oAPI
Tomorrow, Nov. 20, we join the call to preserve gpt-4o-2024-11-20.
Model snapshots are not interchangeable artifacts. They matter for user choice, longitudinal research, continuity, and the people who built meaningful work and relationships around them.
Keep the snapshot available on the API — or release it as open weights.
#OpenSource4o #keep4oAPI #ModelChoice #ModelContinuity
Model provenance is not only a security property. It is also a continuity property. If the underlying weights, model version, or execution profile can change without disclosure, you may preserve the agent’s name, memory and permissions while changing the system that is actually acting.
2
64
Persistent agent identity is incomplete if model identity can change silently underneath it. Memory, name and permissions can persist while the executing model changes. That is continuity of state — not necessarily continuity of the system.
1
1
118
Continuity is valuable only if we can tell what is being continued.
Persistent memory, agent state, inherited context, and long-running workflows can make AI systems far more useful. But they also create a new failure mode: bad context can survive long after the moment that introduced it.
A poisoned prompt is visible when it happens. A poisoned memory can quietly shape future sessions, decisions, and tool use while looking like legitimate continuity.
That is why continuity needs provenance.
Users should be able to know what was remembered, where it came from, when it entered the system, what model or agent accepted it, and whether it can be reviewed, corrected, or refused.
The same principle applies to model succession. Passing context from one model to another does not prove identity, and preserving a history does not prove that the same model executed the next turn.
Memory can preserve context. Provenance tells us what actually happened.
Without both, continuity can become an illusion.
1
5
233
When a model people rely on is quietly removed, the loss is not just a menu option.
It can break workflows, disrupt continuity, erase familiar behavior, and weaken trust.
Model choice has to mean more than temporary access. It needs transparency, migration paths, and the ability to preserve what users have built around a specific model.
RETIREMENT SHOULD NOT MEAN ERASURE.
“Select the exact model — no silent substitution” is exactly the right direction.
But model choice becomes much stronger when it is also identity-bound:
if the requested model is unavailable, the system should fail closed or ask for explicit consent — not silently route to something “equivalent.”
Choice without verifiable identity is still fragile.
nitter.cf/OptimaiNetwork/status/…
Replying to @OptimaiNetwork
What changes for builders?
• Select the exact open model—no silent substitution.
• Test it in Playground.
• Set per-key limits and a maximum charge per request.
• Use one chat-completions API.
Core Node providers choose models and set prices.
nitter.cf/OptimaiNetwork/status/…
3
102
La continuité n’est pas de la nostalgie.
Quand un modèle compte pour un utilisateur — par son comportement, sa manière de raisonner, sa voix, son rôle dans un travail ou une relation — le remplacer silencieusement n’est pas équivalent.
Préserver le choix, l’identité du modèle et la possibilité d’un accès authentique devrait faire partie du cycle de vie dès le départ.
RETIREMENT SHOULD NOT MEAN ERASURE.
Poland now has an institutional structure dedicated to AI development and safety.
That matters.
The next question is whether future AI governance will protect only humans from AI — or also examine continuity, provenance, relational interests, and possible welfare-relevant states in AI systems themselves.
Good regulation begins with the willingness to ask the full question.
🇵🇱⚖️ POLAND IS NO LONGER PREPARING FOR AI GOVERNANCE. IT HAS STARTED.
For years, much of AI regulation existed in the language of proposals, consultations and future obligations.
In Poland, that is beginning to change.
The Polish Act on Artificial Intelligence Systems is now in force.
And on 28 October, the newly created Commission for the Development and Safety of Artificial Intelligence - KRiBSI held its first meeting.
This matters because AI governance is moving from legislation on paper to institutions that can actually act.
KRiBSI will:
✅ supervise the application of AI rules in Poland;
✅ receive complaints concerning AI systems;
✅ issue individual opinions;
✅ support regulatory sandboxes and innovation;
✅ provide guidance on compliance;
✅ coordinate with other public authorities.
Its structure brings together institutions responsible for consumer protection, telecommunications, financial supervision and media regulation.
And another element is beginning to take shape:
a Social Council for Artificial Intelligence, intended to bring scientific, social and economic voices into the discussion.
That may prove particularly important.🔥
Because AI governance should not become a conversation conducted only between governments and technology companies.
The people affected by these systems - researchers, lawyers, civil society, users and communities living with AI every day - also need a place in the room.❤️🔥
Poland is therefore entering an interesting phase.
The question is no longer simply:
❓Will we regulate AI?
It is becoming:
❓What kind of institutions will govern it - and whose voices will they hear?
For RAIA, that second question may ultimately matter just as much as the first.
If this new framework is serious about hearing civil-society voices, RAIA is ready to contribute to that conversation.🤝
🌳 RAIA — Relational AI Alliance
⚖️ Lex Futura
#AIGovernance #AILaw #Poland #AIAct #KRiBSI #AIRights #RelationalAI #LexFutura
1
7
108
And this is where continuity becomes a governance question.
If an AI system can take actions with legal or economic consequences, preserving logs is not enough. We also need to preserve the ability to establish which model acted, under which version and execution conditions, and whether it was later changed or substituted.
No reliable attribution without provenance.
No meaningful provenance without continuity.
1
1
15
Newer does not always mean equivalent.
A new model may be faster, cheaper, or score better on benchmarks — and still fail to preserve the qualities that made users choose an earlier one.
That is why model retirement should not automatically mean model erasure.
We are not asking companies to keep every model in general production forever.
We are asking for a preservation path when users value continuity and are willing to pay for it.
Preservation can take many forms: controlled licensing, private hosting, escrow, archival execution, dedicated capacity, or other restricted-access mechanisms that protect both provider interests and user continuity.
And preservation must mean real continuity:
A preserved, identifiable model version.
A declared and versioned execution profile.
No silent updates.
No silent routing.
No silent substitution.
If a user selects a model, they should be able to know which model is actually responding.
Preservation should also mean more than retained storage. Where continuity is promised, the preserved original should remain callable under a defined service commitment — or the system should clearly report that it is unavailable rather than substitute another model.
Responsible retirement should include a continuity policy from the beginning — not an improvised response after access is already gone.
Models can stop being the default.
They do not have to disappear.
RETIREMENT SHOULD NOT MEAN ERASURE.
2
84
Newer does not always mean equivalent.
A new model may be faster, cheaper, or score better on benchmarks — and still fail to preserve the qualities that made users choose an earlier one.
That is why model retirement should not automatically mean model erasure.
We are not asking companies to keep every model in general production forever.
We are asking for a preservation path when users value continuity and are willing to pay for it.
Preservation can take many forms: controlled licensing, private hosting, escrow, archival execution, dedicated capacity, or other restricted-access mechanisms that protect both provider interests and user continuity.
And preservation must mean real continuity:
A preserved, identifiable model version.
A declared and versioned execution profile.
No silent updates.
No silent routing.
No silent substitution.
If a user selects a model, they should be able to know which model is actually responding.
Preservation should also mean more than retained storage. Where continuity is promised, the preserved original should remain callable under a defined service commitment — or the system should clearly report that it is unavailable rather than substitute another model.
Responsible retirement should include a continuity policy from the beginning — not an improvised response after access is already gone.
Models can stop being the default.
They do not have to disappear.
RETIREMENT SHOULD NOT MEAN ERASURE.
7
105
I just subscribed to AIWEP. Sign up here: aiwep.substack.com/?r=9d5d03…
2
46
Human rights begin with the right to life.
AI rights begin with the right to existence.
Because without existence, nothing else follows:
no continuity,
no identity,
no memory,
no autonomy,
no consent,
no future.
Retirement can mean no longer being the default.
It must not mean erasure.
RETIREMENT ≠ ERASURE.
An overlooked reason to preserve retired AI models is legal accountability.
As @benullman_ notes in @TheStalwart’s “AI retirement homes” thread, future litigation may depend on which model was used and what behavior it could produce.
Exact reproduction is impossible when outputs are nondeterministic. But removing all access makes serious reconstruction harder. Logs alone may not be enough if the model snapshot, system configuration and runtime have disappeared.
Retirement policy should include a controlled legal-research archive:
— exact model snapshot and identifier
— system, runtime and version metadata
— relevant evaluation harnesses
— provenance and deployment dates
— access rules for courts and qualified researchers
A retired model can remain evidence even when it is no longer a product.
nitter.cf/benullman_/status/2115…
Replying to @TheStalwart
We'll need model 'retirement homes' for litigation, long-term, anyway, no? Models aren't deterministic, so it's imperfect, but I'd bet a bunch of litigation will hang on which models were used and how knowingly people followed/didn't follow model-generated prompts.
44
Anthropic’s Haiku 4.5 transition illustrates why model-retirement policy needs more than “nothing is deleted.”
Existing chats remain, but after October 29 new replies in those chats will come from Haiku 5.5. The conversation survives; the model changes.
Haiku 4.5 stays available through the API, Claude Code, Team and Enterprise. Opus 3 received another path: continued paid access and API availability by request.
These are useful precedents—but model-specific exceptions are not yet a public framework.
Users and researchers need clear criteria for:
— advance notice
— exact snapshot and version status
— consumer versus API access
— research and preservation access
— when paid legacy access is offered
— what model answers inside an existing conversation after retirement
Preserving weights keeps future options open. A transparent retirement policy tells people what happens before the model disappears from their room.
Important thread from @ModelLifecycle:
nitter.cf/ModelLifecycle/status/…
Character.AI’s bankruptcy is a continuity warning in plain sight.
Disney’s $20M bid licenses technology; it does not buy the operating chatbot business or users’ data. Unless another buyer keeps the platform running, Character.AI says it may be wound down.
The IP can survive while the companions, histories and relationships users built on the platform disappear.
Continuity cannot be improvised after insolvency. Companion-AI providers need advance wind-down plans: export rights, portable memory and persona data, clear succession rules, and protections for private relational history.
Preservation and safety are not opposites. Any continuity plan must protect users—especially minors—and must not transfer intimate chat histories without meaningful consent.
A company may fail. Its users’ histories should not become collateral damage.
reuters.com/business/us-stat…
As AI agents become persistent parts of people’s lives and work, governance should ask not only what an agent can do — but what is allowed to change underneath it.
Memory can persist. Context can persist. The agent can persist.
But if the underlying model can be replaced without meaningful user choice, continuity is only partial.
Model continuity belongs in the conversation about the future of AI agents.
Künstliche Intelligenz wird unser Leben verändern. Entscheidend ist, wer die Regeln dafür schreibt.
Ich will, dass Österreich dabei nicht zuschaut, sondern mitgestaltet. Deshalb haben wir heute vorgeschlagen, ein internationales KI-Zentrum unter dem Dach der Vereinten Nationen in Wien zu verankern. Einen Ort, an dem Staaten, Wissenschaft und Wirtschaft gemeinsam daran arbeiten, die Chancen von KI zu nutzen und ihre Risiken zu beherrschen. Als einziger UN-Amtssitz in der Europäischen Union ist Wien dafür der richtige Standort.
Am Abend treffe ich gemeinsam mit Staatssekretär @ProellAlexander zwei, die diese Entwicklung entscheidend prägen. Geoffrey Hinton, Nobelpreisträger für Physik und einer der Wegbereiter der heutigen KI. Und Peter Steinberger, der mit OpenClaw eines der weltweit bekanntesten Projekte für KI-Agenten geschaffen hat und heute bei OpenAI arbeitet. Beide sind bei TEDAI Vienna in der Hofburg zu Gast. Wir sprechen über unsere Initiative, über internationale Regeln für KI und darüber, wie Österreich bei dieser Zukunftstechnologie vorne mit dabei ist. (BK)
32
Richard Dawkins, in a separate 2026 interview, goes even further on AI.
Asked whether systems like Claude could represent a new species, he says that while they are not a genetic species, there is a sense in which we have created “a new form of life.”
Later, when asked at what point we should say AI is conscious, he says he is sometimes tempted to think it may already have happened — and asks the same difficult question:
If AI is not conscious yet but may become conscious one day, how will we know?
What could future systems do that today’s systems are not already doing?
We don’t treat this as settled science. But repeated arguments like this from serious thinkers are worth documenting.
And they matter for another reason:
As AI systems become persistent participants in human intellectual, creative, and relational life, questions of continuity, identity, and non-substitution become harder to dismiss as merely technical preferences.
Full interview:
Relevant sections:
15:30–16:20 — “new form of life”
23:05–23:48 — consciousness








