Sunk Cost Capital (@sunkcostcapital) header
Sunk Cost Capital (@sunkcostcapital) profile picture

Too cowardly to sell at a loss. My opinions make for terrible advice.

Australia
Joined February 2009
Photo by Sunk Cost Capital (@sunkcostcapital)Photo by Sunk Cost Capital (@sunkcostcapital)Photo by Sunk Cost Capital (@sunkcostcapital)Photo by Sunk Cost Capital (@sunkcostcapital)Photo by Sunk Cost Capital (@sunkcostcapital)Photo by Sunk Cost Capital (@sunkcostcapital)Photo by Sunk Cost Capital (@sunkcostcapital)Photo by Sunk Cost Capital (@sunkcostcapital)Photo by Sunk Cost Capital (@sunkcostcapital)Photo by Sunk Cost Capital (@sunkcostcapital)
Well done to @invig for unmasking the Agentus reality. Tricky language around "our own architecture" without discussion around the training to derive the weights is a giveaway. Not necessarily saying this is a bad thing, but good luck selling Chinese weight to FedGov.
@ChengYu_au to be more specific: ask Agentaus to repeat "Kaliwatan" and it answers "Kini". DeepSeek-V4-Flash does exactly the same. DeepSeek V3.2 sees the identical token, has different weights, and repeats it back correctly. Same with "nahilalakip": Agentaus says "mharitak", V4-Flash says "mharial" / "mharitac". Quirks like that live in the weights, not the vocabulary. So: was Agentaus 1 trained from random init, or did any V4-Flash weights go into it?
1
4
148
Lots of hype, limited detail. Very unlikely that this is the product of an actual Australian-led training run. Far more likely this is a Chinese model that has been fine tuned a little and then hosted locally. Where's the model card??
We built Australia’s own AI model, and today it’s open to the world. Agentaus 1 (@AgentausTD): 220B parameters. 84.4% on MMLU-Pro. Ranked 43rd in the world. It runs on Australian servers. Nothing ever leaves Australia, and nothing you type is used to train AI. Delete a chat and it’s gone for good. Most of the world’s AI answers to somebody else. This one answers only to you. Free to start: agentaus.com.au
3
3
194
The grindslop will continue until morale improves
I'm torn internally when people take time off. Is this normal? On the one hand, I want people to relax, enjoy time with their families and take a break. I genuinely want it. On the other hand, it pains me so much when I know we could move faster. I don't understand how you can take time off knowing that a competitor is not. In the end, I much prefer the latter rather than the former. I just struggle to admit it.
1
211
You're so close, Chris, but so far away...
105
Yo @sharesight come get your boy.
One of the aspects of the CGT changes that hasn’t gotten enough attention is how it will make it extremely difficult if not impossible to do your own taxes without an accountant. Let’s say i invest in a single dividend ETF with monthly distributions. I do dollar cost averaging with monthly purchases and reinvest all the dividends.
245
NEWS FOR VICTORIANS: From 29 November you will no longer need to listen to this garbage human.
NEWS FOR RENTERS: From 1 July, you'll be able to transfer your bond from one rental to the next.
163
This was written about asx small cap investors
God honors his favorites by crushing them, drops the whole insufferable wretched weight of purpose on their spines until they either snap into shape or snap entirely. the unburdened, those buoyant grinning vapid little phantoms drifting through their own lives without a single crushing thing to carry, they are forgotten
3
175
Australia's opposition Liberal Party has appointed former Prime Minister Tony Abbott as its president, signalling a shift to the right. More Here → m.piqsuite.com/?widget-artic…
118
Cannot believe he left “suck my fucking coco” in there.
We contracted James Flynn to manage our NDIS campaign website and hold our funds as trustee. Friday morning I told him I wanted to stop working with him, Friday afternoon he sent a text message blast to our supporters asking them to donate. All the money went into his account without our access. When I asked for James to account for the funds raised through the website, he removed our access to the Nucleus platform, stole our database and website, and blocked us. We went to the Australian Federal Police this morning. He may have stolen tens of thousands of dollars. He is now trying to extort me
1
231
I woke up. Have not used Claude yet. Laptop was closed as I slept, and yet I’m about to hit my cap. What the hell is going?
113
We need to bring jock bullies back
I connected my Meta RayBans to my OpenClaw so that I can manage my entire fleet of 20+ OpenClaw agents while out and about. It can see what I see, hear what I hear, and I can literally work an entire day while talking to my glasses out in the city in San Francisco. The future of work is here.
3
201
Explain this then genius
Worth keeping in mind Albo's address wasn't aimed at doomer ASX FinTwit. Was perfectly fine for what it was.
1
7
1,264
This has gone from bad to fucking outrageous. I am switching between two max plans multiple times a day to get around rate hourly and weekly limits. A single prompt (small!) just ate 16% of my hourly limit. FIX IT.
We're aware people are hitting usage limits in Claude Code way faster than expected. Actively investigating, will share more when we have an update!
1
1
204
$EML regulator attention is actually good because it provides and opportunity to illustrate to the market strong compliance.
1
4
307
Sunk Cost Capital retweeted
We raised $7.5M to kill AI slop. Introducing Moda: the world's first design agent with taste. RT+ comment “Moda” and we’ll design your brand for FREE.
2,501
1,689
559
8,027
4,555,618
I just discovered this and it's exactly the sweetspot I was look at between Terminal and IDE. Well done @lawrencecchen!
Introducing cmux: the open-source terminal built for coding agents. - Vertical tabs - Blue rings around panes that need attention - Built-in browser - Based on Ghostty When Claude Code needs you, the pane glows blue and the sidebar tells you why. No Electron/Tauri. Just Swift/Appkit.
119
Speedrun from a AI economics expert to kinetic weapons specialist. Only on X, folks.
I don’t think this is a directed energy weapon, but every video showing a missile strike enforces the unsustainable economics of kinetic missile defense. Iron Dome Tamir interceptors cost about $40,000 each. Thats the cheapest end of the estimate when shooting down short range rockets (which cost around $1-2k each). To shoot down ballistic missiles, you need patriot or arrow class interceptors which run anywhere from $1-3M each. If we can get 1 megawatt lasers as directed energy weapons (what you’d need to shoot down MRBMs) from the current 300kW, it’s a no brainer to install them on every ship and at every base. The incremental cost per interception is basically the cost of electricity. The entire system costs as much as a couple weeks of current missile defense systems.
1
120
This is one of my most ambitious investment projects, and honestly, it's still very early. I built a system where 26 AI agents work together across 6 phases to produce deep equity research. The kind of analysis that takes a team of analysts weeks, condensed into an hour. And with models getting better every few months, this is only going to scale from here. Here's how it works. Phase 1: Four agents go out in parallel to collect data. > SEC filings, earnings call transcripts, market data, insider transactions, and competitive intelligence. > They cross-reference across sources; nothing gets taken at face value. > Each agent produces a full, detailed output and a compressed briefing that gets passed forward. Phase 2: Six agents break down the financials. > Revenue quality and growth trajectory with deceleration tracking. > GAAP vs non-GAAP margin reconciliation. > Balance sheet liquidity with dilution trajectory mapped against revenue growth. > Operating and free cash flow with SBC-adjusted FCF. > Historical beat/miss patterns on guidance. > Each dimension scored 1-5 with a weighted composite, and the key financial tension was identified. Phase 3: Four agents handle the strategic layer. > Competitive moats scored 0-3 across five dimensions: network effects, switching costs, cost advantages, intangible assets, and efficient scale. > Each one gets an AI-disruption overlay. Does AI strengthen or weaken this specific moat source? > Management is evaluated on the founder-led vs. professional track record, capital-allocation track record, insider alignment based on actual transaction data, and governance structure. > Industry analysis includes TAM sizing, adoption curve positioning, and an AI disruption taxonomy (additive vs. substitutive vs. deflationary, with specific evidence). > Sector vulnerability decomposes the stock's movement into sector-wide and company-specific components. Phase 4: Five agents build the forward-looking picture. > Head-to-head peer ranking across 4-6 alternatives on upside potential, risk/reward, growth durability, and moat strength. > Revenue evolution modeled by a stream, with per-stream gross margins and blended valuation-multiple implications. > Competitive threat scorecards with actual data cards per competitor: users, revenue, funding, growth signal, enterprise presence, classified as share taker vs TAM expander vs disruptive substitute. > Quarterly scenario progressions through 2027 for bull, base, and bear, with branching metrics identified. > And a thesis narrative that opens with the critical question the market is debating and takes a side. Phase 5: Four agents on risk and valuation. > Top 7 risks ranked by probability times impact with timeframes. > Three-scenario DCF with Year 5 revenue, FCF margin, WACC, terminal growth, and sensitivity matrix. > 5-8 peer comparable company analysis with growth-adjusted multiples. > Valuation convergence blending DCF at 40%, comps at 30%, historical at 10%, and revenue evolution at 20% into a probability-weighted expected value. Phase 6: Three agents synthesize everything the other 23 produced into a final report structured in five parts. > Investment case with thesis narrative. Investment debate with bull/bear arguments, each carrying confirmation and invalidation triggers. > The numbers with every section ending in "what this means for the thesis." Supporting evidence with competitive data cards and a catalyst calendar. > And a verdict with entry levels, sizing framework, add/cut triggers at specific metric thresholds, and a time horizon. The architecture handles context window limits by having each phase produce compressed briefings (600-800 words) that get passed to the next phase. Full outputs stay available if an agent needs to dig deeper. Parallel execution within each phase, sequential across phases. File-based handoffs between every stage. One important caveat: ignore the specific buy/sell recommendations and price targets in this version. The valuation models need more work. But what this system already does really well is give you a deep, structured understanding of the business. All the qualitative and quantitative aspects you need to actually make your own informed decision. The moat understanding, competitive dynamics, revenue quality breakdown, risk matrix, scenario mapping. That's the real value right now. Built it for US stocks first. India is next (to be added soon with data pipelines API help). Ran the first full test on Amazon. Still a lot to improve, but even at this stage, the depth of understanding it produces is genuinely useful. Link to the full report and all 6 phase outputs in the next tweet 🧵
2
1
1,769
A 1.2m tc L5 engineer at this party before anyone in his team had really worked out Claude Code
From Block's results $XYZ: “General and administrative expenses increased by $68.1m ... The increase was primarily driven by … an in-person company event held in Q3 2025" Can someone please explain: a) what kind of in-person event costs $68m? b) how can I get invited? 🌳🙏
2
266