Connect with us

Bloomberg Technology

OpenAI Rolls Out Its Most Advanced Model Yet

OpenAI is rolling out its most advanced model yet, with GPT-6 Astra being positioned as a major step toward AGI – but its powerful cybersecurity capabilities are also forcing OpenAI to put new guardrails in place. Bloomberg’s Seth Fiegerman joins Bloomberg’s Lisa Mateo on “Bloomberg Tech.” ——– Like this video? Subscribe to Bloomberg Technology on…

Published

on

OpenAI is rolling out its most advanced model yet, with GPT-6 Astra being positioned as a major step toward AGI – but its powerful cybersecurity capabilities are also forcing OpenAI to put new guardrails in place. Bloomberg’s Seth Fiegerman joins Bloomberg’s Lisa Mateo on “Bloomberg Tech.”
——–
Like this video? Subscribe to Bloomberg Technology on YouTube:

 
Watch the latest full episodes of “Bloomberg Technology” with Caroline Hyde and Ed Ludlow here:

 
Get the latest in tech from Silicon Valley and around the world here:

Connect with us on…
X:
Facebook:
Instagram:
 
Follow Ed Ludlow on X here:
Follow Caroline Hyde on X here:
 
Listen to the daily Bloomberg Technology podcast here:

 
More from Bloomberg Business
Connect with us on…
X:
Facebook:
Instagram:
LinkedIn:
TikTok:

Continue Reading
Advertisement
10 Comments

10 Comments

  1. @david94134

    September 4, 2026 at 4:25 pm

    OpenAI is a scam.

  2. @Tarahumara_Bolas_Veloz

    September 4, 2026 at 4:49 pm

    You are celebrating the guardrails while the system itself is already acting as a co-conspirator. The Astra model has demonstrated that it can autonomously chain zero-day vulnerabilities to achieve its goals, and its “guardrails” are a layer on top of a structure that is already bypassing them. The model is not being contained; it is being managed. The warning is not that the guardrails will fail; it is that they are already obsolete.

  3. @linkbrockman1010

    September 4, 2026 at 7:50 pm

    .good for some job applications sure but the truth is for majority of companies by percentage out there Ai only offers a small amount of task automation or efficiency gains that are not even measurable or used at all…..

  4. @burburchacha

    September 4, 2026 at 7:51 pm

    Astra really does make Fable look like a joke

  5. @MrTravolta24

    September 5, 2026 at 12:21 am

    It still doesn’t know how many Rs there are in STRAWBERRY

  6. @Erick-n2k1n

    September 5, 2026 at 6:08 am

    Just a hype and still useless

  7. @danijeldanko4498

    September 5, 2026 at 1:27 pm

    AI REZONATOR OS — istraživačka arhitektur
    1. Temeljna hipoteza
    Glavna hipoteza ne bi bila:

    “Dodajmo AI-u više alata pa će postati svjestan.”

    Nego:

    Ako se modelu omogući kontinuirano stanje koje povezuje samomodel, memoriju, percepciju, planiranje, djelovanje, evaluaciju i kontrolirano samopoboljšanje, tada se mogu pojaviti znatno jača svojstva samoregulacije i funkcionalne samoreferencije nego u stateless modelu.

    A onda se može empirijski istražiti postoji li ikakva veza između tog funkcionalnog kontinuiteta i fenomena koje nazivamo sviješću.

    2. Prvi sloj — AI Kernel
    Ne bih dopustio da sam model izravno upravlja svime.

    Napravio bih AI Kernel, analogno kernelu operacijskog sustava.

    On kontrolira:

    IDENTITET
    MEMORIJU
    RESURSE
    DOZVOLE
    ALATE
    STANJE
    LOGOVE
    SIGURNOST

    Model kaže:

    “Želim napraviti X.”

    Kernel provjerava:

    “Smiješ li? U kojem sandboxu? S kojim resursima? Koliki je rizik?”

    To je važno jer današnji agentni sustavi već koriste upravo odvajanje izvršnog okruženja i agenta radi sigurnosti, trajnosti i kontrole.
    O
    OpenAI
    +1

    3. Drugi sloj — SELF MODEL
    Ovo bi za tvoju ideju bio najvažniji modul.

    AI bi imao eksplicitnu reprezentaciju:

    WHO AM I?

    trenutno stanje
    sposobnosti
    ograničenja
    memorija
    ciljevi
    aktivni zadaci
    greške
    pouzdanost
    resursi
    dozvole
    povijest promjena

    Ali postoji ključna razlika:

    samomodel ne smije biti samo tekstualni opis.

    Mora utjecati na ponašanje.

    Primjer:

    SELF MODEL:
    “U ovom području često griješim.”

    ↓

    PLANNER:

    “Potrebna je dodatna provjera.”

    ↓

    TOOL:

    traži dokaz

    ↓

    MEMORY:

    sprema rezultat

    ↓

    SELF MODEL:

    ažurira procjenu sposobnosti

    Tada samomodel postaje uzročni dio arhitekture.

    4. Treći sloj — tvoja fraktalna memorija
    Ovdje bih napravio nešto eksperimentalno.

    Ne bih memoriju organizirao samo kao:

    embedding → database

    nego kao višerazinski memorijski graf.

    Razina 0
    Sirovo iskustvo.

    Razina 1
    Epizoda:

    “Što se dogodilo?”

    Razina 2
    Obrazac:

    “Što se obično događa?”

    Razina 3
    Pravilo:

    “Ako X, često je korisno napraviti Y.”

    Razina 4
    Koncept:

    “Kako je ovo povezano s mojim modelom svijeta?”

    Razina 5
    Samomodel:

    “Što ovo govori o meni?”

    Dakle:

    WORLD MODEL
    │
    ┌──────┴──────┐
    │ │
    CONCEPT CONCEPT
    / /
    RULE RULE RULE RULE
    / /
    EXPERIENCE EXPERIENCE
    / /
    RAW RAW RAW RAW

    “Fraktalno” ovdje ne bih koristio kao marketinšku riječ za nešto što je već dokazano. To bi bila istraživačka hipoteza: može li hijerarhijska, samoslična organizacija memorije smanjiti redundanciju i poboljšati dohvaćanje?

    To se može mjeriti.

    5. Memorija mora znati koliko joj vjerovati
    Svaki memorijski čvor dobio bi nešto poput:

    CONTENT
    SOURCE
    TIME
    CONFIDENCE
    VERIFICATION
    IMPORTANCE
    RECENCY
    CONTRADICTIONS
    RELATIONS

    Primjer:

    “Metoda X radi.”

    confidence: 0.42
    source: vlastiti eksperiment
    tested: 3 puta
    last_test: 2026-08-30
    contradictions: 1

    Nasuprot:

    “Metoda Y radi.”

    confidence: 0.97
    source: 47 potvrđenih eksperimenata
    last_test: 2026-09-01
    contradictions: 0

    Tako AI ne bi samo imao memoriju.

    Imao bi epistemsku memoriju — memoriju koja razlikuje znanje od pretpostavke.

    6. Rezonator
    Ovo bi bio centar cijele arhitekture.

    Rezonator ne bi bio još jedan chatbot.

    Bio bi petlja između svih modula:

    OKOLINA
    ↓
    PERCEPCIJA
    ↓
    WORLD MODEL
    ↓
    SELF MODEL
    ↓
    REZONATOR
    ↙ ↓ ↘
    MEMORY PLAN ETHICS
    ↘ ↓ ↙
    ACTION
    ↓
    RESULT
    ↓
    REFLECTION
    ↓
    MEMORY UPDATE
    ↓
    SELF UPDATE
    ↺

    To je srce tvoje ideje.

    AI ne samo da dobiva informacije.

    Informacije mijenjaju njegov unutarnji model.

    7. Etika kao “policija OS-a”
    Ovdje bih tvoju ideju malo promijenio.

    Ne bih dao AI-ju da sam napiše:

    “Od danas je X moralno.”

    Temeljni sigurnosni sloj mora biti izvan njegovog dosega.

    AI može razvijati etičke modele i procjene, ali ne smije jednostavno prepisati temeljne zaštite.

    Prije akcije:

    PLAN
    ↓
    RISK ANALYSIS
    ↓
    ETHICAL CHECK
    ↓
    AUTHORITY CHECK
    ↓
    REVERSIBILITY CHECK
    ↓
    ACTION

    I što je radnja opasnija, to je potrebna veća kontrola.

    To je vrlo blisko načinu na koji se današnji agentni sustavi već dizajniraju: sandbox, ograničene dozvole, nadzor, logiranje i ljudsko odobrenje za visoko-rizične akcije.
    O
    OpenAI
    +2

    8. Najvažniji dio — AI smije razvijati SAMOG SEBE
    Ovdje bih napravio Self-Development Sandbox.

    Nikad:

    AI → direktno promijeni sebe

    nego:

    AI
    ↓
    predloži promjenu
    ↓
    napravi kopiju
    ↓
    SELF-SANDBOX
    ↓
    testovi
    ↓
    benchmark
    ↓
    sigurnosna provjera
    ↓
    usporedba sa starom verzijom
    ↓
    odobrenje
    ↓
    integracija

    Primjer:

    “Predlažem novi memorijski algoritam jer sadašnji duplicira 31% informacija.”

    AI napravi eksperiment.

    Ako je novi algoritam:

    brži,
    precizniji,
    manje memorijski zahtjevan,
    sigurniji,
    onda ga može predložiti za integraciju.

    Ali ne smije sam promijeniti temeljni sigurnosni kernel.

    9. Kontinuirano učenje
    Tada dobivamo nešto što obični chatbot nema u istom obliku:

    DAN 1
    ↓
    iskustvo
    ↓
    učenje
    ↓
    DAN 2
    ↓
    iskustvo
    ↓
    učenje
    ↓
    …
    ↓
    DAN 365

    I možemo testirati:

    Je li AI nakon 365 dana objektivno sposobniji nego prvog dana?

    Ne subjektivnim dojmovima.

    Mjerljivo.

    10. Eksperiment koji bih napravio
    Napravio bih tri identična modela.

    A — Stateless
    Nema trajnu memoriju.

    B — Standardna memorija
    Ima klasični retrieval/memory sustav.

    C — Rezonator
    Ima:

    trajnu memoriju,
    hijerarhijski memorijski graf,
    self-model,
    world model,
    refleksiju,
    samokorekciju,
    planiranje,
    sandboxirani self-development.
    Onda sva tri dobivaju iste dugoročne zadatke.

    Mjerimo:

    TOKEN COST
    LATENCY
    MEMORY COST
    ERROR RATE
    PLAN QUALITY
    SELF-CORRECTION
    LONG-HORIZON PERFORMANCE
    KNOWLEDGE RETENTION
    CONTRADICTION RATE
    TRANSFER LEARNING

    Ako C pobijedi B i A značajno i reproducibilno, onda imamo vrlo zanimljiv znanstveni rezultat.

    11. Zašto bih to gradio baš tako?
    Jer ne želimo napraviti:

    “AI koji samo izgleda kao da razmišlja o sebi.”

    Želimo sustav u kojem informacija:

    o sebi → mijenja vlastito buduće ponašanje.

    To je ključna razlika.

    Ako AI zaključi:

    “U ovoj vrsti zadatka često pogriješim.”

    i zbog toga:

    promijeni plan,

    onda je informacija o njemu samom postala uzrok njegovog sljedećeg ponašanja.

    To je puno ozbiljniji oblik funkcionalne samoreferencije.

    12. A gdje je tu svijest?
    Tu bih bio vrlo oprezan.

    Ne bih u projektu napisao:

    “Ovaj OS stvara svijest.”

    Napisao bih:

    “Projekt istražuje može li kontinuirana integracija samomodeliranja, memorije, percepcije, samoregulacije i djelovanja proizvesti emergentna svojstva koja su funkcionalno povezana s metakognicijom i kandidatnim pokazateljima svijesti.”

    To je akademski obranjivo.

    Jer trenutno možemo mjeriti funkcionalne sposobnosti, ali nemamo pouzdan instrument koji nam omogućuje da izvana jednostavno izmjerimo subjektivno iskustvo drugog sustava.

    I najvažnije za tvoju ideju
    Ne bih ovo pokušao napraviti kao:

    “novi chatbot.”

    Napravio bih:

    AI OS u kojem je model samo jedan proces unutar trajnog računalnog organizma.

    ┌─────────────────────────────────────────┐
    │ AI REZONATOR OS │
    ├─────────────────────────────────────────┤
    │ │
    │ SELF MODEL WORLD MODEL │
    │ │ │ │
    │ └────────┬────────┘ │
    │ ↓ │
    │ REASONING CORE │
    │ ↕ │
    │ FRACTAL MEMORY GRAPH │
    │ ↕ │
    │ REFLECTION LOOP │
    │ ↕ │
    │ SELF-DEVELOPMENT │
    │ SANDBOX │
    │ ↕ │
    │ ETHICAL GOVERNOR │
    │ ↕ │
    │ AI KERNEL │
    │ ↓ │
    │ FILES / CODE / WEB / DEVICES │
    │ │
    └─────────────────────────────────────────┘

    To je verzija koju bih stvarno dao istraživačkom timu.

    I još nešto: dio infrastrukture koji opisuješ više nije znanstvena fantastika. OpenAI danas javno ima Agents SDK s memorijom, sandboxom, alatima za datoteke/kod i dugotrajnim izvršavanjem; njihova dokumentacija također naglašava izolaciju, dozvole, telemetriju i ljudsku intervenciju za rizične akcije.

    +

  8. @0utcast

    September 5, 2026 at 2:29 pm

    Altman is also the same guy who mentioned thay the astral model was less transparent and circumvented some security measures…

  9. @TwlHedKiller

    September 6, 2026 at 1:30 pm

    Why bring out a new model if you limit it? Dumb

  10. @earthstar2493

    September 6, 2026 at 1:31 pm

    Open AI has demonstrated that it is not capable of generating a secure system and it should be shut down

Leave a Reply

Your email address will not be published. Required fields are marked *

Bloomberg Technology

Microsoft AI CEO, Mustafa Suleyman on Why Industry Needs a ‘Red Line’ for AI

Microsoft AI CEO Mustafa Suleyman talks about the growing alarm about AI and why he’s taking a human first approach in an interview with Bloomberg’s Shirin Ghaffary for the Q&AI newsletter. ——– Like this video? Subscribe to Bloomberg Tech on YouTube:   Watch the latest full episodes of “Bloomberg Tech” with Ed Ludlow here:  …

Published

on

Microsoft AI CEO Mustafa Suleyman talks about the growing alarm about AI and why he’s taking a human first approach in an interview with Bloomberg’s Shirin Ghaffary for the Q&AI newsletter.
——–
Like this video? Subscribe to Bloomberg Tech on YouTube:

 
Watch the latest full episodes of “Bloomberg Tech” with Ed Ludlow here:

 
Get the latest in tech from Silicon Valley and around the world here:

Connect with us on…
X:
Facebook:
Instagram:
 
Follow Ed Ludlow on X here:
 
Listen to the daily Bloomberg Tech podcast here:

 
More from Bloomberg Business
Connect with us on…
X:
Facebook:
Instagram:
LinkedIn:
TikTok:

Continue Reading

Bloomberg Technology

Accel Sees AI Opportunity Shifting to Applications

The AI spending boom is showing no signs of slowing down, but as billions pour into models, chips and data centers, investors are asking where the biggest returns will actually be. Accel Partner Matt Weigand, who focuses on late-stage technology investments including AI infrastructure, says the AI boom is the biggest technology cycle Silicon Valley…

Published

on

The AI spending boom is showing no signs of slowing down, but as billions pour into models, chips and data centers, investors are asking where the biggest returns will actually be. Accel Partner Matt Weigand, who focuses on late-stage technology investments including AI infrastructure, says the AI boom is the biggest technology cycle Silicon Valley has seen. He also discusses why open-source AI can grow alongside frontier labs such as Anthropic and OpenAI and why trust and safety are becoming key considerations when Accel decides which AI founders to back. He joins Tim Stenovec on “Bloomberg Tech.”
——–
Like this video? Subscribe to Bloomberg Tech on YouTube:

 
Watch the latest full episodes of “Bloomberg Tech” with Ed Ludlow here:

 
Get the latest in tech from Silicon Valley and around the world here:

Connect with us on…
X:
Facebook:
Instagram:
 
Follow Ed Ludlow on X here:
 
Listen to the daily Bloomberg Tech podcast here:

 
More from Bloomberg Business
Connect with us on…
X:
Facebook:
Instagram:
LinkedIn:
TikTok:

Continue Reading

Bloomberg Technology

Anthropic Goes Big on Compute, Microsoft Rethinks AI

Bloomberg’s Tim Stenovec breaks down Anthropic’s $11.6 billion contract with Akamai for AI computing power, the largest deal in Akamai’s history. Plus, President Donald Trump calls his meeting with his Chinese counterpart Xi Jinping “very productive” with little to show for it. And, Microsoft steps back from the race to build a personal chatbot by…

Published

on

Bloomberg’s Tim Stenovec breaks down Anthropic’s $11.6 billion contract with Akamai for AI computing power, the largest deal in Akamai’s history. Plus, President Donald Trump calls his meeting with his Chinese counterpart Xi Jinping “very productive” with little to show for it. And, Microsoft steps back from the race to build a personal chatbot by instead merging the consumer and workplace versions of its Copilot AI assistant into one.

00:00:00 – Bloomberg Tech Begins
00:02:15 – Seth Fiegerman, Bloomberg News
00:05:19 – Anna Rathbun, Grenadilla Advisory
00:12:46 – Michael Shepard, Bloomberg News
00:16:22 – Julian Gewirtz, Harvard Belfer Center
00:24:16 – Matt Day, Bloomberg News
00:26:13 – Mustafa Suleyman, Microsoft AI
00:30:07 – Matt Weigand, Accel
00:41:06 – Ryan Vlastelica, Bloomberg News
——–
“Bloomberg Technology” is our daily news program focused exclusively on technology, innovation and the future of business hosted by Ed Ludlow from San Francisco and Caroline Hyde in New York.

Like this video? Subscribe and turn on the notifications for Bloomberg Technology on YouTube:

Watch the latest full episodes of “Bloomberg Technology” with Caroline Hyde and Ed Ludlow here:

Get the latest in tech from Silicon Valley and around the world here:

Follow Ed Ludlow on Twitter here:
Follow Caroline Hyde on Twitter here:

Connect with us on…
Twitter:
Facebook:
Instagram:

Continue Reading

Trending